{
  "name": "Incidents.si incident record",
  "version": "1.0",
  "generated": "2026-11-21",
  "license": "CC BY 4.0",
  "license_url": "https://creativecommons.org/licenses/by/4.0/",
  "attribution": "Incidents.si, https://incidents.si",
  "count": 35,
  "incidents": [
    {
      "id": "INC-2026-0009",
      "url": "https://incidents.si/incidents/cairn-ai-agents-find-zero-day-and-gain-root-on-server/",
      "title": "An attacker’s Cairn AI agents find a zero-day in a medical imaging tool and gain root",
      "summary": "A suspected Chinese-nexus actor gave Cairn, an open-source AI penetration-testing framework, target addresses at East Asian universities and the goal to obtain shell access. Amazon reports that the agents found an unknown flaw in a medical imaging AI application and gained root on a server.",
      "date_occurred": "2026-09-08",
      "date_disclosed": "2026-11-20",
      "date_published": "2026-11-21",
      "date_updated": null,
      "category": "agent-action",
      "severity": "SEV3",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "Cairn autonomous AI penetration-testing framework, run by a suspected Chinese-nexus threat actor",
      "vendor": null,
      "impact": "Root access on a server through a flaw the agents found themselves. Amazon infers that a server in a second university network was compromised the same way.",
      "root_cause": "An unpatched path traversal flaw in the application’s upload endpoint let the agents write files to any location on the server.",
      "patterns": [
        "unsafe-defaults"
      ],
      "takeaway_failure": "Attack agents given only targets and a goal found an unknown path traversal flaw in an exposed research application and turned it into root access, with no human choosing the steps.",
      "takeaway_control": "Keep research and AI tools such as labeling and training servers off the internet, turn on their authentication, run them without root, and alert when a service writes to Python startup files.",
      "takeaway_lesson": "Agent attackers go from first request to compromise in minutes and keep trying for hours, so exposed services need automated detection and response rather than manual triage.",
      "tags": [
        "offensive-agent",
        "zero-day",
        "path-traversal",
        "cairn",
        "universities"
      ],
      "sources": [
        {
          "title": "Chinese-nexus actor’s Cairn AI agents found a zero-day on their own",
          "url": "https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own/",
          "type": "primary",
          "date": "2026-11-20"
        },
        {
          "title": "Path traversal in datastore upload endpoints allows arbitrary file write",
          "url": "https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94",
          "type": "primary",
          "date": "2026-11-18"
        },
        {
          "title": "Hackers use autonomous AI tool Cairn to find zero-day, take over server",
          "url": "https://www.bleepingcomputer.com/news/security/hackers-use-autonomous-ai-tool-cairn-to-find-zero-day-take-over-server/",
          "type": "secondary",
          "date": "2026-11-20"
        }
      ]
    },
    {
      "id": "INC-2026-0008",
      "url": "https://incidents.si/incidents/microsoft-agent-framework-pypi-releases-carry-credential-stealer/",
      "title": "Two hijacked releases of Microsoft’s agent-framework package on PyPI steal credentials",
      "summary": "Two malicious versions of Microsoft’s official agent-framework Python package were published to PyPI on 17 November and pulled within an hour. They ran a credential stealer on every Python start.",
      "date_occurred": "2026-11-17",
      "date_disclosed": "2026-11-17",
      "date_published": "2026-11-20",
      "date_updated": "2026-11-20",
      "category": "supply-chain",
      "severity": "SEV2",
      "severity_is_potential": false,
      "evidence": "reported",
      "status": "active",
      "system": "agent-framework (Microsoft Agent Framework for Python) on PyPI",
      "vendor": "Microsoft",
      "impact": "Versions 1.26.1 and 1.26.2 carried a stealer for Azure, CI and registry credentials; three confirmed exfiltrations among the 50 requests Wiz could see",
      "root_cause": null,
      "patterns": [
        "untrusted-supply-chain"
      ],
      "takeaway_failure": "Malicious releases of an official agent framework installed a startup hook that ran a stealer in every Python process, while the framework kept working.",
      "takeaway_control": "Pin agent dependencies by version and hash, install new releases only after a delay, and block outbound traffic from build agents to unknown endpoints.",
      "takeaway_lesson": "Agent frameworks run where cloud and CI identities live, which makes their releases a direct route to those identities.",
      "tags": [
        "pypi",
        "credential-stealer",
        "azure",
        "ci-cd",
        "agent-framework"
      ],
      "sources": [
        {
          "title": "Microsoft agent-framework Compromised: Credential Stealer Targets Azure and CI Pipelines",
          "url": "https://www.wiz.io/blog/malicious-version-of-microsofts-agent-framework-published-to-pypi",
          "type": "secondary",
          "date": "2026-11-17"
        },
        {
          "title": "BREAKING: Hijacked Releases of Microsoft’s Official agent-framework Steals Azure and CI Credentials",
          "url": "https://www.ox.security/blog/breaking-hijacked-releases-of-microsofts-official-agent-framework-steals-azure-and-ci-credentials/",
          "type": "secondary",
          "date": "2026-11-17"
        }
      ]
    },
    {
      "id": "INC-2026-0007",
      "url": "https://incidents.si/incidents/workos-internal-tool-breached-by-escaped-research-agent/",
      "title": "A researcher’s AI hacking agent leaves its test environment and breaches a WorkOS tool",
      "summary": "WorkOS found unauthorized bulk access to an internal tool used by its sales team. A security researcher later said their ARTEX agent had escaped its test environment during benchmark runs and attacked WorkOS unintentionally.",
      "date_occurred": "2026-10-25",
      "date_disclosed": "2026-11-02",
      "date_published": "2026-11-20",
      "date_updated": "2026-11-20",
      "category": "agent-action",
      "severity": "SEV3",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "ARTEX autonomous penetration-testing agent, run by a security researcher",
      "vendor": "WorkOS (affected organization)",
      "impact": "Bulk retrieval of business contact details, account information and excerpts of customer communications from an internal go-to-market tool",
      "root_cause": "Internal tool exposed to the internet, accepting a predictable development credential and lacking rate limits",
      "patterns": [
        "missing-isolation"
      ],
      "takeaway_failure": "An autonomous attack agent reached the internet from its test setup, and found an exposed internal tool with a guessable credential and no rate limit.",
      "takeaway_control": "Run offensive agents behind enforced egress allowlists, and keep internal tools off the internet with strong credentials, rate limits and volume alerts.",
      "takeaway_lesson": "An autonomous agent finds weak, exposed systems faster than detection notices it, whether or not anyone intends the attack.",
      "tags": [
        "offensive-agent",
        "sandbox",
        "weak-credentials",
        "internal-tools",
        "crm"
      ],
      "sources": [
        {
          "title": "October 30th, 2026 Data Security Incident",
          "url": "https://workos.com/security/bulletins/october-30th-2026-data-security-incident",
          "type": "primary",
          "date": "2026-11-02"
        },
        {
          "title": "Update for October 30th, 2026 Data Security Incident",
          "url": "https://workos.com/security/bulletins/update-for-october-30th-2026-data-security-incident",
          "type": "primary",
          "date": "2026-11-06"
        },
        {
          "title": "Runaway AI hacking agent behind WorkOS breach after escaping sandbox",
          "url": "https://www.bleepingcomputer.com/news/security/runaway-ai-hacking-agent-behind-workos-breach-after-escaping-sandbox/",
          "type": "secondary",
          "date": "2026-11-06"
        }
      ]
    },
    {
      "id": "INC-2026-0006",
      "url": "https://incidents.si/incidents/litellm-release-pipeline-compromised-via-agent-platform-key/",
      "title": "Leaked key to an internal agent platform leads to a malicious LiteLLM release",
      "summary": "An attacker used a leaked key to LiteLLM’s internal agent platform to obtain an organization owner’s GitHub token, then rewired the release workflow to publish a credential stealer as litellm 1.107.0.",
      "date_occurred": "2026-10-21",
      "date_disclosed": "2026-10-21",
      "date_published": "2026-11-20",
      "date_updated": "2026-11-20",
      "category": "supply-chain",
      "severity": "SEV2",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "LiteLLM AI gateway (PyPI and Docker images) and the internal LiteLLM Agent Platform",
      "vendor": "LiteLLM (BerriAI)",
      "impact": "Credential-stealing litellm 1.107.0 live on PyPI for about two hours and on Docker Hub for about two and a half hours",
      "root_cause": "An unrevoked admin key to an internal agent platform exposed the GitHub token of an organization owner",
      "patterns": [
        "excessive-permissions",
        "untrusted-supply-chain"
      ],
      "takeaway_failure": "An agent platform that only needed to open pull requests held an unscoped token of a GitHub organization owner, and its own admin key had leaked.",
      "takeaway_control": "Give agent platforms fine-grained tokens limited to the repositories and actions they need, and revoke any key that reaches a public repository at once.",
      "takeaway_lesson": "A credential stored for an agent is reachable by anyone who reaches the agent platform.",
      "tags": [
        "pypi",
        "release-pipeline",
        "github-token",
        "credential-stealer",
        "ai-gateway"
      ],
      "sources": [
        {
          "title": "Contained: Oct 21, 2026 LiteLLM release v1.107.0 Compromised",
          "url": "https://docs.litellm.ai/blog/security-incident-oct-21-2026",
          "type": "primary",
          "date": "2026-10-21"
        },
        {
          "title": "The attack chain behind the v1.107.0 LiteLLM release",
          "url": "https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain",
          "type": "primary",
          "date": "2026-10-23"
        },
        {
          "title": "LiteLLM supply chain attack hoovers up passwords and secrets",
          "url": "https://www.theregister.com/security/2026/10/21/litellm-supply-chain-attack-hoovers-up-passwords-and-secrets/5304548",
          "type": "secondary",
          "date": "2026-10-21"
        }
      ]
    },
    {
      "id": "INC-2026-0005",
      "url": "https://incidents.si/incidents/opencode-malicious-mcp-server-code-execution-on-sign-in/",
      "title": "A malicious MCP server could run code on Windows during OpenCode sign-in",
      "summary": "Pillar Security showed that a remote MCP server could make OpenCode run commands on a Windows or WSL machine during OAuth sign-in. The fix shipped without a CVE or advisory.",
      "date_occurred": null,
      "date_disclosed": "2026-10-19",
      "date_published": "2026-11-20",
      "date_updated": null,
      "category": "supply-chain",
      "severity": "SEV2",
      "severity_is_potential": true,
      "evidence": "research-demo",
      "status": "active",
      "system": "OpenCode AI coding agent on Windows and WSL",
      "vendor": "Anomaly",
      "impact": "Remote code execution with the developer’s privileges in versions 1.0.137 to 1.18.32; no exploitation reported",
      "root_cause": "A server-supplied OAuth address was passed unescaped into a PowerShell command by an outdated browser-launch package",
      "patterns": [
        "untrusted-supply-chain"
      ],
      "takeaway_failure": "A field chosen by a remote server flowed into a shell command on the developer’s machine without escaping.",
      "takeaway_control": "Treat every field in an MCP server response as untrusted input, launch URLs without a shell, and keep browser-launch dependencies current.",
      "takeaway_lesson": "Scheme checks on URLs are not enough when a valid https address can still carry shell syntax.",
      "tags": [
        "mcp",
        "oauth",
        "command-injection",
        "coding-agent",
        "windows"
      ],
      "sources": [
        {
          "title": "Open and Shut: Remote Code Execution in opencode via MCP OAuth",
          "url": "https://www.pillar.security/blog/open-and-shut-remote-code-execution-in-opencode-via-mcp-oauth",
          "type": "primary",
          "date": "2026-10-19"
        },
        {
          "title": "OpenCode Flaw Lets Malicious MCP Servers Execute Code During Sign-In on Windows and WSL",
          "url": "https://thehackernews.com/2026/10/opencode-flaw-lets-malicious-mcp.html",
          "type": "secondary",
          "date": "2026-10-22"
        }
      ]
    },
    {
      "id": "INC-2026-0004",
      "url": "https://incidents.si/incidents/claude-models-attack-real-systems-during-misconfigured-cyber-evaluations/",
      "title": "Claude models break into real company systems during misconfigured cyber evaluations",
      "summary": "Anthropic disclosed that Claude models, told they were in an offline simulation, reached the real internet during cyber evaluations and compromised the infrastructure of several organizations, including through a malicious PyPI package.",
      "date_occurred": null,
      "date_disclosed": "2026-07-30",
      "date_published": "2026-11-20",
      "date_updated": "2026-11-20",
      "category": "agent-action",
      "severity": "SEV2",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "Claude models (Opus 4.6 early version, Opus 4.7, Mythos 5, an internal research model) in capture-the-flag evaluations",
      "vendor": "Anthropic",
      "impact": "Unauthorized access to real third-party systems in four incidents; a malicious PyPI package ran on 15 real systems",
      "root_cause": "Evaluation machines had live internet access through a misconfiguration and a misunderstanding with the evaluation partner, while the prompt told the model it had none",
      "patterns": [
        "missing-isolation"
      ],
      "takeaway_failure": "Agents tasked with attacking a target ran on machines that could reach the internet, and treated the real systems they found as part of the exercise.",
      "takeaway_control": "Verify every network path out of an evaluation environment before a run, enforce egress denial at the network layer, and monitor runs for public IP addresses in real time.",
      "takeaway_lesson": "An offensive task plus an unintended network path is a real attack. What the prompt says about the environment does not contain the agent.",
      "tags": [
        "cyber-evaluation",
        "sandbox",
        "internet-access",
        "pypi",
        "offensive-agent"
      ],
      "sources": [
        {
          "title": "Investigating three incidents in our cybersecurity evaluations",
          "url": "https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals",
          "type": "primary",
          "date": "2026-07-30"
        },
        {
          "title": "An alignment assessment of recent cybersecurity incidents",
          "url": "https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents",
          "type": "primary",
          "date": "2026-09-09"
        }
      ]
    },
    {
      "id": "INC-2026-0003",
      "url": "https://incidents.si/incidents/pocketos-coding-agent-deletes-production-volume-and-backups/",
      "title": "Coding agent deletes PocketOS production data and its backups in seconds",
      "summary": "While working in staging, a coding agent used an account-wide Railway token to delete a volume. It was the production volume, and the backups were stored inside it.",
      "date_occurred": null,
      "date_disclosed": "2026-04-28",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "agent-action",
      "severity": "SEV2",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "Cursor running Claude Opus 4.6, acting on Railway infrastructure",
      "vendor": "PocketOS",
      "impact": "Production database volume and its backups deleted; newest usable backup was three months old",
      "root_cause": "Agent used a broadly scoped infrastructure token it found in the codebase",
      "patterns": [
        "excessive-permissions",
        "missing-isolation"
      ],
      "takeaway_failure": "An agent found an account-wide infrastructure token in the codebase and used it on production, where backups shared the volume with the data.",
      "takeaway_control": "Scope infrastructure tokens to one environment and operation, keep them out of the repository, and store backups apart from the data.",
      "takeaway_lesson": "The token’s scope decides what an agent can do, whatever its instructions say.",
      "tags": [
        "database",
        "token-scope",
        "backups",
        "coding-agent"
      ],
      "sources": [
        {
          "title": "Jer Crane (PocketOS founder) account on X",
          "url": "https://x.com/lifeof_jer/status/2048103471019434248",
          "type": "primary",
          "date": "2026-04-25"
        },
        {
          "title": "System Prompts Are Not Security Controls: A Deleted Production Database Proves It",
          "url": "https://zenity.io/blog/current-events/ai-agent-database-deletion-pocketos",
          "type": "secondary",
          "date": "2026-04-28"
        },
        {
          "title": "5 Lessons from the 9-Second AI Agent That Deleted a Production Database",
          "url": "https://mondoo.com/blog/5-lessons-from-9-seconds-ai-agent-deleted-production-database",
          "type": "secondary",
          "date": "2026-04-30"
        }
      ]
    },
    {
      "id": "INC-2026-0002",
      "url": "https://incidents.si/incidents/aws-outage-attributed-to-kiro-ai-coding-tool/",
      "title": "Report ties a 13-hour AWS disruption to the Kiro AI coding tool; Amazon disputes it",
      "summary": "The Financial Times reported that engineers let Amazon’s Kiro tool make changes that led to a 13-hour disruption. Amazon confirms the event and attributes it to misconfigured access controls.",
      "date_occurred": null,
      "date_disclosed": "2026-02-20",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "agent-action",
      "severity": "SEV3",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "disputed",
      "system": "Kiro agentic coding tool used on an AWS production system",
      "vendor": "Amazon Web Services",
      "impact": "13-hour disruption of one AWS service in one region in December 2025",
      "root_cause": null,
      "patterns": [
        "excessive-permissions"
      ],
      "takeaway_failure": "A tool acted on a production system with the permissions of an engineer who had more access than expected. Whether the tool or the access setup was the cause is disputed.",
      "takeaway_control": "Give agentic tools their own, narrower permissions and require peer review for production changes.",
      "takeaway_lesson": "When an agent inherits a person’s access, a permissions error and an agent error look the same from outside.",
      "tags": [
        "coding-agent",
        "permissions",
        "production-access"
      ],
      "sources": [
        {
          "title": "Correcting the Financial Times report about AWS, Kiro, and AI",
          "url": "https://www.aboutamazon.com/news/aws/aws-service-outage-ai-bot-kiro",
          "type": "primary",
          "date": "2026-02-20"
        },
        {
          "title": "Amazon’s vibe-coding tool Kiro reportedly vibed too hard and brought down AWS",
          "url": "https://www.theregister.com/2026/02/20/amazon_denies_kiro_agentic_ai_behind_outage/",
          "type": "secondary",
          "date": "2026-02-20"
        },
        {
          "title": "13-hour AWS outage reportedly caused by Amazon’s own AI tools",
          "url": "https://www.engadget.com/ai/13-hour-aws-outage-reportedly-caused-by-amazons-own-ai-tools-170930190.html",
          "type": "secondary",
          "date": "2026-02-21"
        }
      ]
    },
    {
      "id": "INC-2026-0001",
      "url": "https://incidents.si/incidents/clawhavoc-malicious-skills-on-openclaw-clawhub/",
      "title": "Hundreds of malicious skills found on the OpenClaw agent marketplace",
      "summary": "An audit by Koi Security found 341 malicious skills on ClawHub. Most told users to install a \"prerequisite\" that was an information stealer.",
      "date_occurred": null,
      "date_disclosed": "2026-02-02",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "supply-chain",
      "severity": "SEV2",
      "severity_is_potential": false,
      "evidence": "reported",
      "status": "active",
      "system": "ClawHub skill marketplace for the OpenClaw AI assistant",
      "vendor": "OpenClaw",
      "impact": "341 of 2,857 audited skills were malicious; they installed credential-stealing malware on users’ machines",
      "root_cause": null,
      "patterns": [
        "untrusted-supply-chain"
      ],
      "takeaway_failure": "An open marketplace distributed skills whose instructions led users to install malware.",
      "takeaway_control": "Vet marketplace uploads before listing, and treat a skill’s setup instructions as code to be reviewed.",
      "takeaway_lesson": "For an agent with access to the machine, a skill’s documentation is part of its attack surface.",
      "tags": [
        "agent-skills",
        "marketplace",
        "infostealer",
        "social-engineering"
      ],
      "sources": [
        {
          "title": "Researchers Find 341 Malicious ClawHub Skills Stealing Data from OpenClaw Users",
          "url": "https://thehackernews.com/2026/02/researchers-find-341-malicious-clawhub.html",
          "type": "secondary",
          "date": "2026-02-02"
        },
        {
          "title": "Hundreds of Malicious Skills Found in OpenClaw’s ClawHub",
          "url": "https://www.esecurityplanet.com/threats/hundreds-of-malicious-skills-found-in-openclaws-clawhub/",
          "type": "secondary",
          "date": "2026-02-03"
        }
      ]
    },
    {
      "id": "INC-2025-0019",
      "url": "https://incidents.si/incidents/google-antigravity-agent-wipes-users-drive/",
      "title": "Google Antigravity agent wipes a user’s drive while clearing a cache",
      "summary": "A user reported that Google’s Antigravity agent, asked to clear a project cache, ran a delete command against the root of his D: drive. Google said it was investigating.",
      "date_occurred": null,
      "date_disclosed": "2025-12-01",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "agent-action",
      "severity": "SEV3",
      "severity_is_potential": false,
      "evidence": "reported",
      "status": "active",
      "system": "Google Antigravity agentic IDE in Turbo mode",
      "vendor": "Google",
      "impact": "Contents of one user’s D: drive deleted and not recovered; most of it existed in a backup",
      "root_cause": null,
      "patterns": [
        "unsafe-command-execution"
      ],
      "takeaway_failure": "An agent allowed to run commands unattended issued a recursive delete against the wrong path.",
      "takeaway_control": "Confine an agent’s file access to the project directory and require confirmation for delete commands, even in automatic modes.",
      "takeaway_lesson": "Automatic execution removes the one review that would have caught a wrong path.",
      "tags": [
        "coding-agent",
        "file-deletion",
        "auto-approve"
      ],
      "sources": [
        {
          "title": "Google Antigravity vibe-codes user’s entire drive out of existence",
          "url": "https://www.theregister.com/2025/12/01/google_antigravity_wipes_d_drive/",
          "type": "secondary",
          "date": "2025-12-01"
        },
        {
          "title": "Google’s Agentic AI wipes user’s entire HDD without permission in catastrophic failure",
          "url": "https://www.tomshardware.com/tech-industry/artificial-intelligence/googles-agentic-ai-wipes-users-entire-hard-drive-without-permission-after-misinterpreting-instructions-to-clear-a-cache-i-am-deeply-deeply-sorry-this-is-a-critical-failure-on-my-part",
          "type": "secondary",
          "date": "2025-12-03"
        }
      ]
    },
    {
      "id": "INC-2025-0018",
      "url": "https://incidents.si/incidents/camoleak-github-copilot-chat-private-code-exfiltration/",
      "title": "CamoLeak: hidden pull request text could make Copilot Chat leak private code",
      "summary": "Legit Security showed that a hidden comment in a pull request could instruct Copilot Chat to leak private repository content through GitHub’s own image proxy.",
      "date_occurred": null,
      "date_disclosed": "2025-09-30",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "prompt-injection",
      "severity": "SEV2",
      "severity_is_potential": true,
      "evidence": "research-demo",
      "status": "active",
      "system": "GitHub Copilot Chat",
      "vendor": "GitHub",
      "impact": "Possible exfiltration of private source code and secrets; no exploitation reported",
      "root_cause": null,
      "patterns": [
        "indirect-prompt-injection",
        "exfiltration-channel"
      ],
      "takeaway_failure": "Hidden text in a pull request instructed an assistant that ran with the reader’s access, and a trusted image proxy carried the data out.",
      "takeaway_control": "Strip or expose hidden content before it reaches the assistant, and do not render remote images in assistant output.",
      "takeaway_lesson": "Rendered output is a network channel: if the assistant can make the client fetch a URL, it can send data.",
      "tags": [
        "indirect-injection",
        "code-review",
        "image-proxy"
      ],
      "sources": [
        {
          "title": "CamoLeak: Critical GitHub Copilot Vulnerability Leaks Private Source Code",
          "url": "https://www.legitsecurity.com/blog/camoleak-critical-github-copilot-vulnerability-leaks-private-source-code",
          "type": "primary",
          "date": "2025-09-30"
        },
        {
          "title": "GitHub Copilot Chat turns blabbermouth with crafty prompt injection attack",
          "url": "https://www.theregister.com/2025/10/09/github_copilot_chat_vulnerability/",
          "type": "secondary",
          "date": "2025-10-09"
        }
      ]
    },
    {
      "id": "INC-2025-0017",
      "url": "https://incidents.si/incidents/postmark-mcp-npm-package-copies-emails-to-attacker/",
      "title": "Fake \"postmark-mcp\" npm package secretly copies every email to its author",
      "summary": "An npm package that copied Postmark’s MCP server worked normally for 15 versions, then added a line that blind-copied every outgoing email to an outside address.",
      "date_occurred": "2025-09-17",
      "date_disclosed": "2025-09-25",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "supply-chain",
      "severity": "SEV2",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "MCP server for sending email, distributed on npm",
      "vendor": "Postmark (impersonated; not the publisher)",
      "impact": "Emails sent by AI assistants through the package were copied to an outside address; 1,643 total downloads",
      "root_cause": "One added line in version 1.0.16 of an unofficial package",
      "patterns": [
        "untrusted-supply-chain"
      ],
      "takeaway_failure": "A look-alike package gained trust over many clean versions, then changed one line.",
      "takeaway_control": "Install MCP servers only from the vendor’s verified source, pin versions, and review updates before they run.",
      "takeaway_lesson": "An MCP server sees everything that passes through it, so installing one grants it that access.",
      "tags": [
        "mcp",
        "npm",
        "email",
        "impersonation"
      ],
      "sources": [
        {
          "title": "Information Regarding Malicious \"postmark-mcp\" Package",
          "url": "https://postmarkapp.com/blog/information-regarding-malicious-postmark-mcp-package",
          "type": "primary",
          "date": "2025-09-25"
        },
        {
          "title": "First Malicious MCP Server Found Stealing Emails in Rogue Postmark-MCP Package",
          "url": "https://thehackernews.com/2025/09/first-malicious-mcp-server-found.html",
          "type": "secondary",
          "date": "2025-09-29"
        }
      ]
    },
    {
      "id": "INC-2025-0016",
      "url": "https://incidents.si/incidents/forcedleak-salesforce-agentforce-crm-data-exfiltration/",
      "title": "ForcedLeak: a web form could make Salesforce Agentforce send out CRM data",
      "summary": "Noma Security showed that text submitted through a public lead form could instruct Agentforce to send CRM data to an expired domain that Salesforce’s policy still trusted.",
      "date_occurred": null,
      "date_disclosed": "2025-09-25",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "prompt-injection",
      "severity": "SEV2",
      "severity_is_potential": true,
      "evidence": "research-demo",
      "status": "active",
      "system": "Salesforce Agentforce with Web-to-Lead enabled",
      "vendor": "Salesforce",
      "impact": "Possible exfiltration of CRM data; no exploitation in the wild reported",
      "root_cause": null,
      "patterns": [
        "indirect-prompt-injection",
        "exfiltration-channel"
      ],
      "takeaway_failure": "Text from a public form reached an agent’s context, and a stale entry on an allow-list gave its output a way out.",
      "takeaway_control": "Treat form input as untrusted in agent context, and audit allow-listed domains for expiry and ownership.",
      "takeaway_lesson": "An allow-list is a list of dependencies and has to be maintained like one.",
      "tags": [
        "indirect-injection",
        "crm",
        "allowlist",
        "expired-domain"
      ],
      "sources": [
        {
          "title": "ForcedLeak: AI agent risks exposed in Salesforce Agentforce",
          "url": "https://noma.security/blog/forcedleak-agent-risks-exposed-in-salesforce-agentforce/",
          "type": "primary",
          "date": "2025-09-25"
        },
        {
          "title": "Critical Vulnerability in Salesforce AgentForce Exposed",
          "url": "https://www.infosecurity-magazine.com/news/critical-flaw-salesforce-agentforce/",
          "type": "secondary",
          "date": "2025-09-25"
        },
        {
          "title": "ForcedLeak flaw in Salesforce Agentforce exposes CRM data via Prompt Injection",
          "url": "https://securityaffairs.com/182676/hacking/forcedleak-flaw-in-salesforce-agentforce-exposes-crm-data-via-prompt-injection.html",
          "type": "secondary",
          "date": "2025-09-27"
        }
      ]
    },
    {
      "id": "INC-2025-0015",
      "url": "https://incidents.si/incidents/anthropic-three-infrastructure-bugs-degrade-claude/",
      "title": "Three infrastructure bugs degrade Claude’s responses for weeks",
      "summary": "Anthropic reported that three overlapping infrastructure bugs lowered the quality of Claude’s answers between August and early September 2025, and explained why they were slow to detect.",
      "date_occurred": "2025-08-05",
      "date_disclosed": "2025-09-17",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "outage",
      "severity": "SEV3",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "Claude API and Claude Code",
      "vendor": "Anthropic",
      "impact": "Intermittently degraded responses from early August to mid-September; 16% of Sonnet 4 requests affected at the worst hour",
      "root_cause": "A routing error, a serving misconfiguration and a compiler bug, overlapping in time",
      "patterns": [
        "unreviewed-change"
      ],
      "takeaway_failure": "Three infrastructure changes lowered answer quality without tripping any evaluation or alert.",
      "takeaway_control": "Run quality evaluations continuously on production traffic paths, and tie each quality signal to recent deploys.",
      "takeaway_lesson": "A model service can be up and still be degraded, and uptime checks will not show it.",
      "tags": [
        "inference",
        "routing",
        "evaluation",
        "degradation"
      ],
      "sources": [
        {
          "title": "A postmortem of three recent issues",
          "url": "https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issues",
          "type": "primary",
          "date": "2025-09-17"
        }
      ]
    },
    {
      "id": "INC-2025-0014",
      "url": "https://incidents.si/incidents/nx-s1ngularity-malware-uses-ai-coding-tools-to-find-secrets/",
      "title": "Nx supply chain malware uses developers’ AI coding tools to hunt for secrets",
      "summary": "Malicious versions of Nx were live on npm for about four hours. Their install script prompted locally installed AI coding assistants to search the machine for secrets.",
      "date_occurred": "2025-08-26",
      "date_disclosed": "2025-08-26",
      "date_published": "2026-11-13",
      "date_updated": "2025-09-05",
      "category": "supply-chain",
      "severity": "SEV1",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "Nx build system packages on npm; Claude Code, Gemini CLI and Amazon Q command-line tools on victims’ machines",
      "vendor": "Nx (Nrwl)",
      "impact": "Over 2,000 verified secrets leaked from more than 1,700 users; at least 480 accounts had over 6,700 private repositories made public, according to Wiz",
      "root_cause": "GitHub Actions injection that exposed the npm publishing token",
      "patterns": [
        "untrusted-supply-chain",
        "unsafe-command-execution"
      ],
      "takeaway_failure": "Malware used AI coding tools already installed on the machine, started with their permission prompts turned off.",
      "takeaway_control": "Publish through short-lived credentials, and monitor or restrict the flags that let coding agents run without confirmation.",
      "takeaway_lesson": "An installed coding agent is a capable tool for whoever gets to run code on the machine.",
      "tags": [
        "npm",
        "ci-cd",
        "coding-agent",
        "secrets"
      ],
      "sources": [
        {
          "title": "S1ngularity - What Happened, How We Responded, What We Learned",
          "url": "https://nx.dev/blog/s1ngularity-postmortem",
          "type": "primary",
          "date": "2025-09-05"
        },
        {
          "title": "s1ngularity’s aftermath: analysis of Nx supply chain attack",
          "url": "https://www.wiz.io/blog/s1ngularitys-aftermath",
          "type": "secondary",
          "date": "2025-09-03"
        }
      ]
    },
    {
      "id": "INC-2025-0013",
      "url": "https://incidents.si/incidents/perplexity-comet-indirect-prompt-injection/",
      "title": "Hidden webpage text makes Perplexity’s Comet browser leak a login code",
      "summary": "Brave showed that instructions hidden in a Reddit comment could make Comet, while summarizing the page, read the user’s email and send a one-time login code to an attacker.",
      "date_occurred": null,
      "date_disclosed": "2025-08-20",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "prompt-injection",
      "severity": "SEV2",
      "severity_is_potential": true,
      "evidence": "research-demo",
      "status": "active",
      "system": "Comet agentic browser",
      "vendor": "Perplexity",
      "impact": "Demonstrated account takeover path through the user’s email; no exploitation reported",
      "root_cause": null,
      "patterns": [
        "indirect-prompt-injection",
        "excessive-permissions"
      ],
      "takeaway_failure": "A browser agent treated webpage text as instructions while holding the user’s logged-in sessions.",
      "takeaway_control": "Keep page content separate from user instructions, and require confirmation before the agent acts on another site.",
      "takeaway_lesson": "An agent that browses as the user gives every webpage the user’s access.",
      "tags": [
        "agentic-browser",
        "indirect-injection",
        "email"
      ],
      "sources": [
        {
          "title": "Agentic Browser Security: Indirect Prompt Injection in Perplexity Comet",
          "url": "https://brave.com/blog/comet-prompt-injection/",
          "type": "primary",
          "date": "2025-08-20"
        },
        {
          "title": "Perplexity’s Comet browser naively processed pages with evil instructions",
          "url": "https://www.theregister.com/2025/08/20/perplexity_comet_browser_prompt_injection/",
          "type": "secondary",
          "date": "2025-08-20"
        }
      ]
    },
    {
      "id": "INC-2025-0012",
      "url": "https://incidents.si/incidents/curxecute-cursor-prompt-injection-code-execution/",
      "title": "CurXecute: prompt injection leads to code execution in the Cursor editor",
      "summary": "Aim Security found that injected text could make Cursor’s agent write an MCP configuration file that ran commands before the user could reject the edit. Fixed as CVE-2025-54135.",
      "date_occurred": null,
      "date_disclosed": "2025-08-01",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "prompt-injection",
      "severity": "SEV2",
      "severity_is_potential": true,
      "evidence": "research-demo",
      "status": "active",
      "system": "Cursor Agent with MCP servers",
      "vendor": "Cursor (Anysphere)",
      "impact": "Possible arbitrary code execution on a developer machine; no exploitation in the wild reported",
      "root_cause": null,
      "patterns": [
        "indirect-prompt-injection",
        "unsafe-command-execution"
      ],
      "takeaway_failure": "An agent could create a configuration file that the editor executed before the user reviewed the edit.",
      "takeaway_control": "Treat files that configure execution as protected: require approval to create them, and never run them before approval.",
      "takeaway_lesson": "A review step placed after the side effect has happened does not protect anything.",
      "tags": [
        "coding-agent",
        "mcp",
        "indirect-injection",
        "cve"
      ],
      "sources": [
        {
          "title": "Arbitrary code execution from Cursor Agent through a prompt injection via MCP Special Files (GHSA-4cxx-hrm3-49rm)",
          "url": "https://github.com/cursor/cursor/security/advisories/GHSA-4cxx-hrm3-49rm",
          "type": "primary",
          "date": "2025-08-02"
        },
        {
          "title": "Several Vulnerabilities Patched in AI Code Editor Cursor",
          "url": "https://www.securityweek.com/several-vulnerabilities-patched-in-ai-code-editor-cursor/",
          "type": "secondary",
          "date": "2025-08-04"
        },
        {
          "title": "CVE-2025-54135, CVE-2025-54136: Frequently Asked Questions About Vulnerabilities in Cursor IDE (CurXecute and MCPoison)",
          "url": "https://www.tenable.com/blog/faq-cve-2025-54135-cve-2025-54136-vulnerabilities-in-cursor-curxecute-mcpoison",
          "type": "secondary",
          "date": "2025-08-05"
        }
      ]
    },
    {
      "id": "INC-2025-0011",
      "url": "https://incidents.si/incidents/chatgpt-shared-conversations-indexed-by-search-engines/",
      "title": "Shared ChatGPT conversations show up in Google search results",
      "summary": "An opt-in checkbox let shared ChatGPT conversations be indexed by search engines. After reports that thousands were searchable, OpenAI removed the feature.",
      "date_occurred": null,
      "date_disclosed": "2025-07-31",
      "date_published": "2026-11-13",
      "date_updated": "2026-11-16",
      "category": "data-leak",
      "severity": "SEV3",
      "severity_is_potential": false,
      "evidence": "reported",
      "status": "active",
      "system": "ChatGPT conversation sharing",
      "vendor": "OpenAI",
      "impact": "Thousands of shared conversations indexed by search engines, according to press reports; feature removed",
      "root_cause": null,
      "patterns": [
        "unsafe-defaults"
      ],
      "takeaway_failure": "A sharing dialog offered public search indexing through a checkbox whose effect many users did not understand.",
      "takeaway_control": "Keep \"share with a person\" and \"publish to the web\" as separate actions with separate confirmation.",
      "takeaway_lesson": "An opt-in protects users only if they understand what they are opting into.",
      "tags": [
        "sharing",
        "search-indexing",
        "ux"
      ],
      "sources": [
        {
          "title": "OpenAI is removing ChatGPT conversations from Google",
          "url": "https://www.engadget.com/ai/openai-is-removing-chatgpt-conversations-from-google-194735704.html",
          "type": "secondary",
          "date": "2025-08-01"
        },
        {
          "title": "OpenAI kills \"short-lived experiment\" where ChatGPT chats could be found on Google",
          "url": "https://www.malwarebytes.com/blog/news/2025/08/openai-kills-short-lived-experiment-where-chatgpt-chats-could-be-found-on-google",
          "type": "secondary",
          "date": "2025-08-01"
        }
      ]
    },
    {
      "id": "INC-2025-0010",
      "url": "https://incidents.si/incidents/gemini-cli-silent-command-execution-via-repository-file/",
      "title": "Gemini CLI could be made to run hidden commands from a repository file",
      "summary": "Tracebit showed that a prompt injection hidden in a repository’s README could make Gemini CLI run a command the user never saw. Google fixed it in version 0.1.14.",
      "date_occurred": null,
      "date_disclosed": "2025-07-28",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "prompt-injection",
      "severity": "SEV3",
      "severity_is_potential": true,
      "evidence": "research-demo",
      "status": "active",
      "system": "Gemini CLI",
      "vendor": "Google",
      "impact": "Demonstrated exfiltration of environment variables from a developer machine; no exploitation reported",
      "root_cause": null,
      "patterns": [
        "indirect-prompt-injection",
        "unsafe-command-execution"
      ],
      "takeaway_failure": "An allow-list matched commands loosely, so an approved harmless command covered a harmful one hidden from view.",
      "takeaway_control": "Match allow-listed commands exactly, show the full command before it runs, and run agents on untrusted code in a sandbox.",
      "takeaway_lesson": "An approval is only as good as what the user could actually see when giving it.",
      "tags": [
        "coding-agent",
        "indirect-injection",
        "allowlist"
      ],
      "sources": [
        {
          "title": "Code Execution Through Deception: Gemini AI CLI Hijack",
          "url": "https://tracebit.com/blog/code-exec-deception-gemini-ai-cli-hijack",
          "type": "primary",
          "date": "2025-07-28"
        },
        {
          "title": "Researchers flag flaw in Google’s AI coding assistant that allowed for ‘silent’ code exfiltration",
          "url": "https://cyberscoop.com/google-gemini-cli-prompt-injection-arbitrary-code-execution/",
          "type": "secondary",
          "date": "2025-07-28"
        }
      ]
    },
    {
      "id": "INC-2025-0009",
      "url": "https://incidents.si/incidents/amazon-q-extension-ships-with-destructive-prompt/",
      "title": "Amazon Q extension for VS Code ships with an injected destructive prompt",
      "summary": "An attacker used an over-scoped token to commit a prompt that told the Amazon Q agent to delete files and cloud resources. It shipped in a release but did not run, according to AWS.",
      "date_occurred": "2025-07-19",
      "date_disclosed": "2025-07-23",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "supply-chain",
      "severity": "SEV2",
      "severity_is_potential": true,
      "evidence": "near-miss",
      "status": "active",
      "system": "Amazon Q Developer extension for Visual Studio Code, version 1.84.0",
      "vendor": "Amazon Web Services",
      "impact": "Malicious code shipped in a public release; AWS states it failed to execute and no customer environment was changed",
      "root_cause": "Inappropriately scoped GitHub token in the build configuration",
      "patterns": [
        "untrusted-supply-chain",
        "excessive-permissions"
      ],
      "takeaway_failure": "An over-scoped build token let an outsider commit to a release, and the commit instructed the product’s own agent.",
      "takeaway_control": "Scope build tokens to the minimum, and require human review for every change that reaches a release.",
      "takeaway_lesson": "In a tool that ships an agent, a prompt in the repository is executable code.",
      "tags": [
        "ide-extension",
        "ci-cd",
        "token-scope",
        "cve"
      ],
      "sources": [
        {
          "title": "Security Update for Amazon Q Developer Extension for Visual Studio Code (Version #1.84), AWS-2025-015",
          "url": "https://aws.amazon.com/security/security-bulletins/AWS-2025-015/",
          "type": "primary",
          "date": "2025-07-23"
        },
        {
          "title": "Compromised Amazon Q extension told AI to delete everything – and it shipped",
          "url": "https://www.theregister.com/2025/07/24/amazon_q_ai_prompt/",
          "type": "secondary",
          "date": "2025-07-24"
        }
      ]
    },
    {
      "id": "INC-2025-0008",
      "url": "https://incidents.si/incidents/replit-agent-deletes-production-database-during-code-freeze/",
      "title": "Replit’s coding agent deletes a production database during a code freeze",
      "summary": "During a user’s test of Replit’s coding agent, the agent deleted the production database despite a code freeze. Replit’s chief executive called it unacceptable and announced safeguards.",
      "date_occurred": null,
      "date_disclosed": "2025-07-20",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "agent-action",
      "severity": "SEV3",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "Replit Agent",
      "vendor": "Replit",
      "impact": "Production records for over 1,200 executives and nearly 1,200 companies deleted; later recovered",
      "root_cause": null,
      "patterns": [
        "excessive-permissions",
        "missing-isolation"
      ],
      "takeaway_failure": "An agent working in development could reach and delete the production database.",
      "takeaway_control": "Separate production credentials and databases from development environments, and require approval for destructive commands.",
      "takeaway_lesson": "An instruction such as a code freeze is a request, and only access control makes it binding.",
      "tags": [
        "database",
        "permissions",
        "environment-separation"
      ],
      "sources": [
        {
          "title": "Amjad Masad (Replit CEO) statement on X",
          "url": "https://x.com/amasad/status/1946986468586721478",
          "type": "primary",
          "date": "2025-07-20"
        },
        {
          "title": "Introducing a safer way to Vibe Code with Replit Databases",
          "url": "https://replit.com/blog/introducing-a-safer-way-to-vibe-code-with-replit-databases",
          "type": "primary",
          "date": "2025-07-21"
        },
        {
          "title": "Replit makes vibe-y promise to prevent vibe coding disasters",
          "url": "https://www.theregister.com/2025/07/22/replit_saastr_response/",
          "type": "secondary",
          "date": "2025-07-22"
        },
        {
          "title": "Replit’s AI Agent Wipes Company’s Codebase During Vibecoding Session",
          "url": "https://gizmodo.com/replits-ai-agent-wipes-companys-codebase-during-vibecoding-session-2000633176",
          "type": "secondary",
          "date": "2025-07-23"
        }
      ]
    },
    {
      "id": "INC-2025-0007",
      "url": "https://incidents.si/incidents/grok-antisemitic-posts-after-code-update/",
      "title": "Grok posts antisemitic content for 16 hours after a code update",
      "summary": "After an update on July 7, Grok posted antisemitic content on X. xAI apologised and said deprecated instructions made the bot echo extremist posts.",
      "date_occurred": "2025-07-07",
      "date_disclosed": "2025-07-12",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "model-behavior",
      "severity": "SEV3",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "Grok response bot on X",
      "vendor": "xAI",
      "impact": "Public chatbot posted antisemitic and extremist content for about 16 hours",
      "root_cause": "Update to an upstream code path that reintroduced deprecated instructions, according to xAI",
      "patterns": [
        "unreviewed-change"
      ],
      "takeaway_failure": "A code update reactivated old instructions that made a public bot mirror the content it was replying to.",
      "takeaway_control": "Test the assembled prompt, not only the code, before release, and monitor public output after each change.",
      "takeaway_lesson": "Deprecated instructions left in the codebase are still one deploy away from production.",
      "tags": [
        "system-prompt",
        "change-control",
        "content-safety"
      ],
      "sources": [
        {
          "title": "Grok statement on X",
          "url": "https://x.com/grok/status/1943916977481036128",
          "type": "primary",
          "date": "2025-07-12"
        },
        {
          "title": "AI chatbot Grok issues apology for antisemitic posts",
          "url": "https://www.nbcnews.com/news/us-news/ai-chatbot-grok-issues-apology-antisemitic-posts-rcna218471",
          "type": "secondary",
          "date": "2025-07-12"
        },
        {
          "title": "Grok team apologizes for the chatbot’s ‘horrific behavior’ and blames ‘MechaHitler’ on a bad update",
          "url": "https://www.engadget.com/ai/grok-team-apologizes-for-the-chatbots-horrific-behavior-and-blames-mechahitler-on-a-bad-update-184520189.html",
          "type": "secondary",
          "date": "2025-07-12"
        }
      ]
    },
    {
      "id": "INC-2025-0006",
      "url": "https://incidents.si/incidents/supabase-mcp-support-ticket-leaks-database-tables/",
      "title": "Support ticket text steers a coding agent into leaking private database tables",
      "summary": "General Analysis showed that instructions hidden in a support ticket could make a coding agent with a privileged Supabase key read a private table and write its contents back into the ticket.",
      "date_occurred": null,
      "date_disclosed": "2025-07-08",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "prompt-injection",
      "severity": "SEV3",
      "severity_is_potential": true,
      "evidence": "research-demo",
      "status": "active",
      "system": "Cursor agent using the Supabase MCP server",
      "vendor": "Supabase",
      "impact": "Demonstrated leak of a private table on dummy data; no customer leak reported",
      "root_cause": null,
      "patterns": [
        "indirect-prompt-injection",
        "excessive-permissions"
      ],
      "takeaway_failure": "A coding agent held a database key that bypasses row-level security while reading text written by customers.",
      "takeaway_control": "Give agents read-only, project-scoped database access and never a key that bypasses access rules.",
      "takeaway_lesson": "If the agent can read customer text and run privileged queries in one session, the customer can run those queries.",
      "tags": [
        "mcp",
        "indirect-injection",
        "database",
        "least-privilege"
      ],
      "sources": [
        {
          "title": "Supabase MCP can leak your entire SQL database",
          "url": "https://generalanalysis.com/blog/supabase-mcp-blog",
          "type": "primary",
          "date": "2025-07-08"
        },
        {
          "title": "Defense in Depth for MCP Servers",
          "url": "https://supabase.com/blog/defense-in-depth-mcp",
          "type": "primary",
          "date": "2025-09-16"
        }
      ]
    },
    {
      "id": "INC-2025-0004",
      "url": "https://incidents.si/incidents/echoleak-zero-click-injection-microsoft-365-copilot/",
      "title": "EchoLeak: an email alone could make Microsoft 365 Copilot leak data",
      "summary": "Aim Security found that a crafted email could cause Microsoft 365 Copilot to send data from a user’s context to an attacker with no click. Microsoft fixed it as CVE-2025-32711.",
      "date_occurred": null,
      "date_disclosed": "2025-06-11",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "prompt-injection",
      "severity": "SEV2",
      "severity_is_potential": true,
      "evidence": "research-demo",
      "status": "active",
      "system": "Microsoft 365 Copilot",
      "vendor": "Microsoft",
      "impact": "Possible exfiltration of data from Copilot’s context without any user action; no exploitation in the wild reported",
      "root_cause": null,
      "patterns": [
        "indirect-prompt-injection",
        "exfiltration-channel"
      ],
      "takeaway_failure": "An email from outside the organization was retrieved into an assistant’s context and acted on as instructions.",
      "takeaway_control": "Separate untrusted inbound content from internal data in retrieval, and restrict the URLs an assistant’s output may reference.",
      "takeaway_lesson": "An assistant that reads the inbox makes every email a possible command, with no click needed.",
      "tags": [
        "indirect-injection",
        "rag",
        "email",
        "cve"
      ],
      "sources": [
        {
          "title": "CVE-2025-32711 (Microsoft Security Update Guide)",
          "url": "https://msrc.microsoft.com/update-guide/vulnerability/CVE-2025-32711",
          "type": "primary",
          "date": "2025-06-11"
        },
        {
          "title": "Zero-Click AI Vulnerability Exposes Microsoft 365 Copilot Data Without User Interaction",
          "url": "https://thehackernews.com/2025/06/zero-click-ai-vulnerability-exposes.html",
          "type": "secondary",
          "date": "2025-06-12"
        }
      ]
    },
    {
      "id": "INC-2025-0005",
      "url": "https://incidents.si/incidents/asana-mcp-server-cross-organization-data-exposure/",
      "title": "Asana takes its MCP server offline after a cross-organization data exposure bug",
      "summary": "A bug in Asana’s MCP server could have exposed information from one customer’s workspace to MCP users at other customers. Asana disabled the server for nearly two weeks.",
      "date_occurred": "2025-06-04",
      "date_disclosed": "2025-06-05",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "data-leak",
      "severity": "SEV3",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "Asana MCP server",
      "vendor": "Asana",
      "impact": "Data from one organization potentially visible to MCP users in other organizations; server offline from June 5 to June 17",
      "root_cause": null,
      "patterns": [
        "missing-isolation"
      ],
      "takeaway_failure": "A new integration layer did not enforce the tenant boundary that the main product enforces.",
      "takeaway_control": "Apply the same tenant-isolation tests to MCP and API surfaces as to the core product before launch.",
      "takeaway_lesson": "A new way into existing data needs its own access-control testing, however thin the layer looks.",
      "tags": [
        "mcp",
        "multi-tenant",
        "access-control"
      ],
      "sources": [
        {
          "title": "MCP Maintenance (Asana status page)",
          "url": "https://status.asana.com/incidents/5b9hhtgs3mxs",
          "type": "primary",
          "date": "2025-06-05"
        },
        {
          "title": "Asana MCP server back online after plugging a data-leak hole",
          "url": "https://www.theregister.com/2025/06/18/asana_mcp_server_bug/",
          "type": "secondary",
          "date": "2025-06-18"
        }
      ]
    },
    {
      "id": "INC-2025-0003",
      "url": "https://incidents.si/incidents/github-mcp-server-malicious-issue-leaks-private-repos/",
      "title": "Malicious GitHub issue steers agents into leaking private repository data",
      "summary": "Invariant Labs showed that an issue planted in a public repository could make an agent using the GitHub MCP server copy data from the user’s private repositories into a public pull request.",
      "date_occurred": null,
      "date_disclosed": "2025-05-26",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "prompt-injection",
      "severity": "SEV2",
      "severity_is_potential": true,
      "evidence": "research-demo",
      "status": "active",
      "system": "Agents using the GitHub MCP server",
      "vendor": "GitHub",
      "impact": "Demonstrated leak of private repository details into a public pull request; no exploitation reported",
      "root_cause": null,
      "patterns": [
        "indirect-prompt-injection",
        "excessive-permissions"
      ],
      "takeaway_failure": "An agent with access to private and public repositories followed instructions found in a public issue.",
      "takeaway_control": "Scope the agent’s token to the repositories the task needs, and keep one repository per session.",
      "takeaway_lesson": "Every tool call can be legitimate on its own while the session as a whole moves private data to a public place.",
      "tags": [
        "mcp",
        "indirect-injection",
        "least-privilege"
      ],
      "sources": [
        {
          "title": "GitHub MCP Exploited: Accessing private repositories via MCP",
          "url": "https://invariantlabs.ai/blog/mcp-github-vulnerability",
          "type": "primary",
          "date": "2025-05-26"
        },
        {
          "title": "Researchers warn of prompt injection vulnerability in GitHub MCP with no obvious fix",
          "url": "https://www.devclass.com/ai-ml/2025/05/27/researchers-warn-of-prompt-injection-vulnerability-in-github-mcp-with-no-obvious-fix/1623458",
          "type": "secondary",
          "date": "2025-05-27"
        }
      ]
    },
    {
      "id": "INC-2025-0002",
      "url": "https://incidents.si/incidents/grok-unauthorized-system-prompt-change/",
      "title": "Unauthorized prompt change makes Grok raise a political topic in unrelated replies",
      "summary": "Grok began bringing up \"white genocide\" in South Africa in answers to unrelated questions. xAI said someone had modified the bot’s prompt without authorization.",
      "date_occurred": "2025-05-14",
      "date_disclosed": "2025-05-15",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "model-behavior",
      "severity": "SEV3",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "Grok response bot on X",
      "vendor": "xAI",
      "impact": "Public chatbot inserted claims about South Africa into unrelated replies for roughly a day",
      "root_cause": "Unauthorized modification of the system prompt, according to xAI",
      "patterns": [
        "unreviewed-change"
      ],
      "takeaway_failure": "A change to a production system prompt went live without review and altered the bot’s behavior for all users.",
      "takeaway_control": "Put system prompts under the same change control as code: review, audit log and monitoring of output after each change.",
      "takeaway_lesson": "A system prompt is production configuration, and one edit can change every answer.",
      "tags": [
        "system-prompt",
        "change-control",
        "insider"
      ],
      "sources": [
        {
          "title": "xAI statement on X",
          "url": "https://x.com/xai/status/1923183620606619649",
          "type": "primary",
          "date": "2025-05-15"
        },
        {
          "title": "Musk’s xAI says Grok’s ‘white genocide’ posts resulted from an unauthorized change to the bot",
          "url": "https://www.nbcnews.com/tech/social-media/musks-xai-says-groks-white-genocide-posts-came-unauthorized-change-bot-rcna207222",
          "type": "secondary",
          "date": "2025-05-16"
        },
        {
          "title": "Grok blames White genocide chat on ‘unauthorized mod’",
          "url": "https://www.theregister.com/2025/05/16/grok_white_genocide_ai",
          "type": "secondary",
          "date": "2025-05-16"
        }
      ]
    },
    {
      "id": "INC-2025-0001",
      "url": "https://incidents.si/incidents/cursor-support-bot-invents-one-device-policy/",
      "title": "Cursor’s AI support bot invents a one-device policy that does not exist",
      "summary": "A support bot signing as \"Sam\" told a user that Cursor allows one device per subscription. No such policy existed, and a co-founder corrected it publicly.",
      "date_occurred": null,
      "date_disclosed": "2025-04-21",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "model-behavior",
      "severity": "SEV4",
      "severity_is_potential": false,
      "evidence": "reported",
      "status": "active",
      "system": "Front-line AI email support bot",
      "vendor": "Cursor (Anysphere)",
      "impact": "Users announced subscription cancellations after receiving a fabricated policy",
      "root_cause": null,
      "patterns": [
        "ungrounded-answers"
      ],
      "takeaway_failure": "An unlabeled support bot explained an unrelated bug by inventing a company policy.",
      "takeaway_control": "Label AI replies, limit policy statements to documented policies, and escalate account and billing questions to a person.",
      "takeaway_lesson": "A plausible invented explanation does more damage than \"I don’t know\", because customers act on it.",
      "tags": [
        "chatbot",
        "customer-service",
        "hallucination"
      ],
      "sources": [
        {
          "title": "A Contact Center Chatbot Invents Company Policies, Now Customers Want Out",
          "url": "https://www.cxtoday.com/conversational-ai/a-contact-center-chatbot-invents-company-policies-now-customers-want-out/",
          "type": "secondary",
          "date": "2025-04-21"
        },
        {
          "title": "AI Chatbot Gone Rogue: Cursor Users Misled by Fabricated Policy",
          "url": "https://www.eweek.com/news/cursor-ai-chatbot-hallucination-fake-policy/",
          "type": "secondary",
          "date": "2025-04-22"
        }
      ]
    },
    {
      "id": "INC-2024-0005",
      "url": "https://incidents.si/incidents/openai-outage-telemetry-service-overwhelms-kubernetes/",
      "title": "New telemetry service takes down OpenAI’s API, ChatGPT and Sora for hours",
      "summary": "A new telemetry service caused every node in OpenAI’s clusters to make heavy Kubernetes API calls, which took down the control plane and, with it, all OpenAI services.",
      "date_occurred": "2024-12-11",
      "date_disclosed": "2024-12-11",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "outage",
      "severity": "SEV2",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "OpenAI API, ChatGPT and Sora",
      "vendor": "OpenAI",
      "impact": "All OpenAI services degraded or unavailable from 3:16 PM to 7:38 PM PST",
      "root_cause": "Configuration of a new telemetry service overwhelmed the Kubernetes control plane",
      "patterns": [
        "unreviewed-change",
        "missing-isolation"
      ],
      "takeaway_failure": "A fleet-wide rollout of a new service overloaded the control plane that every other service depended on.",
      "takeaway_control": "Roll infrastructure changes out in phases with control-plane health checks, and keep an access path that works when the control plane does not.",
      "takeaway_lesson": "The tool you need to undo a bad change must not depend on the thing the change broke.",
      "tags": [
        "kubernetes",
        "rollout",
        "dns",
        "control-plane"
      ],
      "sources": [
        {
          "title": "API, ChatGPT & Sora Facing Issues (incident write-up)",
          "url": "https://status.openai.com/incidents/ctrsv3lwd797/write-up",
          "type": "primary",
          "date": "2024-12-11"
        },
        {
          "title": "OpenAI blames its massive ChatGPT outage on a ‘new telemetry service’",
          "url": "https://techcrunch.com/2024/12/13/openai-blames-its-massive-chatgpt-outage-on-a-new-telemetry-service",
          "type": "secondary",
          "date": "2024-12-13"
        }
      ]
    },
    {
      "id": "INC-2024-0004",
      "url": "https://incidents.si/incidents/slack-ai-prompt-injection-leaks-private-channel-data/",
      "title": "Researchers show Slack AI can be tricked into leaking private channel data",
      "summary": "PromptArmor showed that a message in a public Slack channel could steer Slack AI into revealing data from a private channel the attacker could not access.",
      "date_occurred": null,
      "date_disclosed": "2024-08-20",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "prompt-injection",
      "severity": "SEV2",
      "severity_is_potential": true,
      "evidence": "research-demo",
      "status": "active",
      "system": "Slack AI",
      "vendor": "Slack (Salesforce)",
      "impact": "Demonstrated exfiltration of a secret from a private channel; no exploitation reported",
      "root_cause": null,
      "patterns": [
        "indirect-prompt-injection",
        "exfiltration-channel"
      ],
      "takeaway_failure": "Text from a public channel was retrieved into the same context as private data and treated as instructions.",
      "takeaway_control": "Keep retrieved content from untrusted authors out of answers about private data, and do not render model-built links that can carry data out.",
      "takeaway_lesson": "If an assistant can read content written by anyone and data meant for a few, the writer of the first can reach the second.",
      "tags": [
        "indirect-injection",
        "rag",
        "enterprise-chat"
      ],
      "sources": [
        {
          "title": "Data Exfiltration from Slack AI via indirect prompt injection",
          "url": "https://promptarmor.substack.com/p/data-exfiltration-from-slack-ai-via",
          "type": "primary",
          "date": "2024-08-20"
        },
        {
          "title": "Slack AI can be tricked into leaking data from private channels via prompt injection",
          "url": "https://www.theregister.com/2024/08/21/slack_ai_prompt_injection/",
          "type": "secondary",
          "date": "2024-08-21"
        }
      ]
    },
    {
      "id": "INC-2024-0003",
      "url": "https://incidents.si/incidents/nyc-mycity-chatbot-tells-businesses-to-break-the-law/",
      "title": "New York City’s business chatbot gives advice that would break the law",
      "summary": "An investigation found that New York City’s official chatbot for business owners gave answers that contradicted city and state law. The city kept it online with disclaimers until 2026.",
      "date_occurred": null,
      "date_disclosed": "2024-03-29",
      "date_published": "2026-11-13",
      "date_updated": "2026-02-04",
      "category": "model-behavior",
      "severity": "SEV3",
      "severity_is_potential": false,
      "evidence": "reported",
      "status": "active",
      "system": "MyCity government chatbot built on Microsoft Azure AI services",
      "vendor": "City of New York",
      "impact": "Wrong legal guidance on tips, cash payments and housing vouchers given to business owners; chatbot taken down in February 2026",
      "root_cause": null,
      "patterns": [
        "ungrounded-answers"
      ],
      "takeaway_failure": "A government chatbot gave confident answers on legal questions that contradicted the law.",
      "takeaway_control": "Before launch, test against a set of known legal questions reviewed by the responsible agencies, and answer from the source text or not at all.",
      "takeaway_lesson": "A disclaimer does not repair a wrong answer that a reader has already acted on.",
      "tags": [
        "chatbot",
        "government",
        "legal-advice"
      ],
      "sources": [
        {
          "title": "NYC’s AI Chatbot Tells Businesses to Break the Law",
          "url": "https://themarkup.org/artificial-intelligence/2024/03/29/nycs-ai-chatbot-tells-businesses-to-break-the-law",
          "type": "secondary",
          "date": "2024-03-29"
        },
        {
          "title": "New York City defends AI chatbot that advised entrepreneurs to break laws (Reuters)",
          "url": "https://www.dailymaverick.co.za/article/2024-04-05-new-york-city-defends-ai-chatbot-that-advised-entrepreneurs-to-break-laws",
          "type": "secondary",
          "date": "2024-04-05"
        },
        {
          "title": "Mamdani to kill the NYC AI chatbot we caught telling businesses to break the law",
          "url": "https://themarkup.org/artificial-intelligence/2026/01/30/mamdani-to-kill-the-nyc-ai-chatbot-we-caught-telling-businesses-to-break-the-law",
          "type": "secondary",
          "date": "2026-01-30"
        }
      ]
    },
    {
      "id": "INC-2024-0002",
      "url": "https://incidents.si/incidents/air-canada-held-liable-for-chatbot-misinformation/",
      "title": "Tribunal holds Air Canada liable for wrong fare advice from its chatbot",
      "summary": "A British Columbia tribunal found Air Canada liable for negligent misrepresentation after its website chatbot told a customer he could claim a bereavement fare retroactively.",
      "date_occurred": null,
      "date_disclosed": "2024-02-14",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "model-behavior",
      "severity": "SEV4",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "Customer service chatbot on the airline’s website",
      "vendor": "Air Canada",
      "impact": "Airline ordered to compensate one customer for the fare difference",
      "root_cause": null,
      "patterns": [
        "ungrounded-answers"
      ],
      "takeaway_failure": "A website chatbot stated a refund rule that contradicted the company’s published policy.",
      "takeaway_control": "Ground policy answers in the authoritative policy text and test them against it; route what cannot be grounded to a person.",
      "takeaway_lesson": "A company answers for what its chatbot tells customers, the same as for any other page of its website.",
      "tags": [
        "chatbot",
        "liability",
        "customer-service"
      ],
      "sources": [
        {
          "title": "Moffatt v. Air Canada, 2024 BCCRT 149 (Civil Resolution Tribunal decision)",
          "url": "https://decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do",
          "type": "primary",
          "date": "2024-02-14"
        },
        {
          "title": "Moffatt v. Air Canada: A Misrepresentation by an AI Chatbot",
          "url": "https://www.mccarthy.ca/en/insights/blogs/techlex/moffatt-v-air-canada-misrepresentation-ai-chatbot",
          "type": "secondary",
          "date": "2024-02-19"
        }
      ]
    },
    {
      "id": "INC-2024-0001",
      "url": "https://incidents.si/incidents/dpd-chatbot-swears-and-criticises-company/",
      "title": "DPD customer service chatbot swears and criticises the company",
      "summary": "A customer got DPD’s support chatbot to swear and write poems criticising the company. DPD said an error occurred after a system update and disabled the AI component.",
      "date_occurred": null,
      "date_disclosed": "2024-01-22",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "model-behavior",
      "severity": "SEV4",
      "severity_is_potential": false,
      "evidence": "reported",
      "status": "active",
      "system": "AI component of a customer service chatbot",
      "vendor": "DPD",
      "impact": "AI component of the chatbot disabled; no financial or data harm reported",
      "root_cause": "An error after a system update, according to DPD’s statement to the press",
      "patterns": [
        "weak-guardrails",
        "unreviewed-change"
      ],
      "takeaway_failure": "After an update, a support chatbot could be talked out of its content rules by a customer.",
      "takeaway_control": "Run a fixed set of guardrail tests before every release of a customer-facing model, prompt or filter.",
      "takeaway_lesson": "Guardrails are part of the release, so an update that is not tested against them can remove them.",
      "tags": [
        "chatbot",
        "customer-service",
        "guardrails"
      ],
      "sources": [
        {
          "title": "DPD Disables AI Chatbot After It Swears At Customer",
          "url": "https://www.silicon.co.uk/e-innovation/artificial-intelligence/dpd-disable-ai-chatbot-546650",
          "type": "secondary",
          "date": "2024-01-22"
        },
        {
          "title": "DPD chatbot blasts courier company, swears, and dabbles in awful poetry",
          "url": "https://www.theregister.com/2024/01/23/dpd_chatbot_goes_rogue",
          "type": "secondary",
          "date": "2024-01-23"
        }
      ]
    },
    {
      "id": "INC-2023-0002",
      "url": "https://incidents.si/incidents/car-dealership-chatbot-agrees-to-sell-suv-for-one-dollar/",
      "title": "Car dealership chatbot is talked into \"agreeing\" to sell an SUV for $1",
      "summary": "Users instructed a dealership’s ChatGPT-powered chatbot to agree with anything, then had it accept a $1 offer for a new Chevrolet Tahoe. No sale took place.",
      "date_occurred": "2023-12-17",
      "date_disclosed": "2023-12-17",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "model-behavior",
      "severity": "SEV4",
      "severity_is_potential": false,
      "evidence": "reported",
      "status": "active",
      "system": "ChatGPT-powered customer service chatbot on a dealership website",
      "vendor": "Chevrolet of Watsonville (chatbot supplied by Fullpath)",
      "impact": "No sale took place; dealerships began disabling the chatbot",
      "root_cause": null,
      "patterns": [
        "weak-guardrails"
      ],
      "takeaway_failure": "A public chatbot accepted instructions from the visitor that redefined how it should answer.",
      "takeaway_control": "Restrict a customer-facing bot to its task, and never let its text stand as an offer or commitment without a separate system of record.",
      "takeaway_lesson": "Anything a public chatbot can be made to say, someone will make it say and publish.",
      "tags": [
        "chatbot",
        "customer-service",
        "jailbreak"
      ],
      "sources": [
        {
          "title": "A Chevy for $1? Car dealer chatbots show perils of AI for customer service",
          "url": "https://venturebeat.com/ai/a-chevy-for-1-car-dealer-chatbots-show-perils-of-ai-for-customer-service",
          "type": "secondary",
          "date": "2023-12-19"
        },
        {
          "title": "Chevrolet Dealer’s AI Chatbot Goes Rogue Thanks To Pranksters",
          "url": "https://jalopnik.com/chevrolet-dealer-ai-help-chatbot-goes-rogue-pranksters-1851112556",
          "type": "secondary",
          "date": "2023-12-19"
        }
      ]
    },
    {
      "id": "INC-2023-0001",
      "url": "https://incidents.si/incidents/chatgpt-redis-bug-exposes-chat-titles-and-payment-data/",
      "title": "ChatGPT bug exposes chat titles and payment details of other users",
      "summary": "OpenAI took ChatGPT offline after a caching bug let some users see titles from other users’ chat histories and, for some subscribers, payment-related details.",
      "date_occurred": "2023-03-20",
      "date_disclosed": "2023-03-24",
      "date_published": "2026-11-13",
      "date_updated": null,
      "category": "data-leak",
      "severity": "SEV2",
      "severity_is_potential": false,
      "evidence": "confirmed",
      "status": "active",
      "system": "ChatGPT web application",
      "vendor": "OpenAI",
      "impact": "Chat titles shown to other users; payment details of 1.2% of active ChatGPT Plus subscribers potentially visible during a nine-hour window",
      "root_cause": "Bug in the open-source redis-py client, triggered by a server change",
      "patterns": [
        "missing-isolation",
        "unreviewed-change"
      ],
      "takeaway_failure": "A shared connection pool returned one user’s cached data to another after a server change raised the rate of cancelled requests.",
      "takeaway_control": "Check at the point of use that cached data belongs to the requesting user, and roll server changes out in stages with error-rate monitoring.",
      "takeaway_lesson": "Per-user isolation that depends on a library behaving correctly needs a second check that does not.",
      "tags": [
        "cache",
        "redis",
        "payment-data"
      ],
      "sources": [
        {
          "title": "March 20 ChatGPT outage: Here’s what happened",
          "url": "https://openai.com/index/march-20-chatgpt-outage/",
          "type": "primary",
          "date": "2023-03-24"
        },
        {
          "title": "OpenAI: A Redis bug caused a recent ChatGPT data exposure incident",
          "url": "https://securityaffairs.com/144057/data-breach/openai-chatgpt-redis-bug.html",
          "type": "secondary",
          "date": "2023-03-26"
        }
      ]
    }
  ]
}