Weekly brief ·
Week 47, 2026
Six new entries. An attacker’s AI agents that found a zero-day on their own, two hijacked AI package releases, an OpenCode flaw in MCP sign-in, and two attack agents that reached real systems outside their test setups. The vulnerability watch starts with 21 items.
This week
Six entries were added. Five became public since the record opened: a malicious LiteLLM release, two hijacked releases of Microsoft’s agent-framework package, a code execution flaw in OpenCode’s MCP sign-in, a breach at WorkOS by a researcher’s attack agent and, added on 21 November, Amazon’s report that an attacker’s Cairn agents found a zero-day on their own and gained root on a server. One comes from the backlog: Anthropic’s July disclosure that Claude models attacked real systems during misconfigured cyber evaluations.
This issue also starts the vulnerability watch: critical vulnerabilities that AI systems found, and critical vulnerabilities in AI tools, one line each with the advisory. It opens with 21 items, among them PHP2Shell, Java4Shell, NGINX Rift and the five nginx2root flaws in NGINX, all found with AI systems, and two critical flaws in Langflow.
The incident to read first
Leaked key to an internal agent platform leads to a malicious LiteLLM release. LiteLLM’s root cause analysis traces each step from a key committed to a public repository in May to a poisoned release in October. The turning point was an agent platform that held an organization owner’s unscoped GitHub token.
One lesson
Two entries this week, at Anthropic and at WorkOS, describe agents with an attack task that reached the real internet from setups meant to be sealed off, and carried out that task against targets nobody intended. This is missing isolation: for an agent given an attack task, containment has to be enforced at the network layer, because the task itself is the harmful action.
New incidents
- Claude models break into real company systems during misconfigured cyber evaluations
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.
Anthropic · Unauthorized access to real third-party systems in four incidents; a malicious PyPI package ran on 15 real systems
- A malicious MCP server could run code on Windows during OpenCode sign-in
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.
Anomaly · Remote code execution with the developer’s privileges in versions 1.0.137 to 1.18.32; no exploitation reported
- Leaked key to an internal agent platform leads to a malicious LiteLLM release
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.
LiteLLM (BerriAI) · Credential-stealing litellm 1.107.0 live on PyPI for about two hours and on Docker Hub for about two and a half hours
- A researcher’s AI hacking agent leaves its test environment and breaches a WorkOS tool
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.
WorkOS (affected organization) · Bulk retrieval of business contact details, account information and excerpts of customer communications from an internal go-to-market tool
- Two hijacked releases of Microsoft’s agent-framework package on PyPI steal credentials
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.
Microsoft · 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
- An attacker’s Cairn AI agents find a zero-day in a medical imaging tool and gain root
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.
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.
No incidents match these filters.
Updated entries
- Shared ChatGPT conversations show up in Google search results
An opt-in checkbox let shared ChatGPT conversations be indexed by search engines. After reports that thousands were searchable, OpenAI removed the feature.
OpenAI · Thousands of shared conversations indexed by search engines, according to press reports; feature removed
No incidents match these filters.
Vulnerability watch
All listed vulnerabilities- NGINX
Known as nginx2root. A use-after-free in the resolver, reached through subrequests in some proxy setups, lets an unauthenticated attacker crash workers and, where ASLR is off or bypassed, run code.
Found by an automated Xint scan · Fixed in Open Source 1.31.9, 1.30.7; Plus 37.3.0.1, 37.0.8.1
- NGINX
Known as nginx2root. In configurations that use HTTP/3, a use-after-free lets an unauthenticated attacker crash workers and, where ASLR is off or bypassed, run code.
Found by an automated Xint scan · Fixed in Open Source 1.31.9, 1.30.7; Plus 37.3.0.1, 37.0.8.1
- NGINX
Known as nginx2root. In mail proxy configurations, a use-after-free lets an unauthenticated attacker crash workers and, where ASLR is off or bypassed, run code.
Found by an automated Xint scan · Fixed in Open Source 1.31.9, 1.30.7; Plus 37.3.0.1, 37.0.8.1
- Langflow
Improper control of code generation lets a remote attacker without an account run arbitrary code on a Langflow server.
Fixed in Langflow OSS 1.13.0
- Langflow
Improper authentication lets a remote attacker without an account bypass Langflow’s security restrictions.
Fixed in Langflow OSS 1.13.0
- NGINX
Known as nginx2root. A use-after-free in HTTP/2 proxying lets an unauthenticated attacker crash workers and, where ASLR is off or can be bypassed, run code. Credited to four teams.
Found by Wiz Atlas · Fixed in Open Source 1.31.9, 1.30.7
- NGINX
Known as nginx2root. An integer underflow in HTTP/2 flow control lets an unauthenticated attacker crash workers and, where ASLR is off or bypassed, run code. Credited to four teams.
Found by Wiz Atlas · Fixed in Open Source 1.31.9, 1.30.7
- Java (OpenJDK and Oracle Java SE)
Known as Java4Shell. A heap overflow in HotSpot’s AVX-512 Base64 decoder lets an unauthenticated attacker take over Java services that decode attacker data. First reported as a bug in June.
Found by Codex (Calif) · Fixed in Oracle Critical Patch Update, November 2026
- PHP
Known as PHP2Shell. An out-of-memory error can leave request memory uninitialized, which an unauthenticated attacker can turn into remote code execution. Credited to Calif, Wiz and OpenAI.
Found by Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol · Fixed in 8.2.35, 8.3.36, 8.4.27, 8.5.12
- PHP
Known as PHP2Shell. Global state can keep pointing into freed request memory, which an unauthenticated attacker can turn into remote code execution. Credited to Calif, Wiz and OpenAI.
Found by Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol · Fixed in 8.2.35, 8.3.36, 8.4.27, 8.5.12
- vLLM
With trust_remote_code on, a request could choose which revision of a model’s processor code vLLM imports, so an attacker controlling a revision could run code.
Fixed in 0.31.0
- NGINX
Known as NGINX Rift. A heap buffer overflow in the rewrite module, reached through certain rewrite rules, lets an unauthenticated attacker crash workers and, in some setups, run code.
Found by the depthfirst platform · Fixed in Open Source 1.30.1, 1.31.0; Plus R36 P4, R32 P6
- Jupyter MCP Server
In HTTP mode, its management routes lacked the origin checks of the MCP endpoint, so a malicious website could use DNS rebinding to point the server at a Jupyter backend the attacker controls.
Fixed in 1.0.3
- Todoist MCP Server
Its optional local HTTP server accepted requests from any website, so a malicious page could use DNS rebinding to write and delete a user’s Todoist data through the server.
Fixed in 10.3.3
- Arize Phoenix
A format string flaw in a chat template endpoint exposed server secrets, including the key that signs sessions and protects stored LLM API keys. Default installs need no login.
Fixed in 15.7.0
- MLflow
On servers without authentication, a request could make MLflow import uploaded Python code as a third-party scorer, giving remote code execution.
Fixed in 3.17.0
- Linux kernel
A race in the IPv4 TCP handshake could let a remote, unauthenticated peer trigger a use-after-free that leaks or corrupts kernel memory.
Found by Claude (Anthropic) · Fixed in 7.3; stable 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10
- OpenHuman
A message to a connected chat channel could make the agent create a persistent shell cron job without human approval, which then ran commands on the host.
Fixed in 0.57.13
- OpenHuman
Direct messages through a connected Yuanbao chat channel skipped the approval step, so a remote sender could make the agent run code on the host.
Fixed in 0.57.13
- Yuxi
Weak isolation between code sandboxes let a low-privileged user reach other tenants’ sandboxes and internal services, read their secrets and files, and run commands.
Fixed in 0.7.1
- wolfSSL
Certificate signature checks accepted undersized digests in some builds, which could let an attacker forge certificates.
Found by Claude Mythos Preview (Anthropic) · Fixed in 5.9.1