Failure pattern

Excessive permissions

The agent, or a credential it could reach, was able to do more than the task needed.

Preventive controls from the record

IncidentControl
INC-2026-0006
LiteLLM (BerriAI)
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.
INC-2026-0003
PocketOS
Scope infrastructure tokens to one environment and operation, keep them out of the repository, and store backups apart from the data.
INC-2026-0002
Amazon Web Services
Give agentic tools their own, narrower permissions and require peer review for production changes.
INC-2025-0013
Perplexity
Keep page content separate from user instructions, and require confirmation before the agent acts on another site.
INC-2025-0009
Amazon Web Services
Scope build tokens to the minimum, and require human review for every change that reaches a release.
INC-2025-0008
Replit
Separate production credentials and databases from development environments, and require approval for destructive commands.
INC-2025-0006
Supabase
Give agents read-only, project-scoped database access and never a key that bypasses access rules.
INC-2025-0003
GitHub
Scope the agent’s token to the repositories the task needs, and keep one repository per session.

Controls are editorial summaries by Incidents.si, not statements by the organizations involved.

8 incidents with this pattern

  1. INC-2026-0006
    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

    SEV2Supply chainConfirmed
  2. INC-2026-0003
    Coding agent deletes PocketOS production data and its backups in seconds

    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.

    PocketOS · Production database volume and its backups deleted; newest usable backup was three months old

    SEV2Agent actionConfirmed
  3. INC-2026-0002
    Report ties a 13-hour AWS disruption to the Kiro AI coding tool; Amazon disputes it

    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.

    Amazon Web Services · 13-hour disruption of one AWS service in one region in December 2025

    SEV3Agent actionConfirmeddisputed
  4. INC-2025-0013
    Hidden webpage text makes Perplexity’s Comet browser leak a login code

    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.

    Perplexity · Demonstrated account takeover path through the user’s email; no exploitation reported

    SEV2potentialPrompt injectionDemonstrated
  5. INC-2025-0009
    Amazon Q extension for VS Code ships with an injected destructive prompt

    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.

    Amazon Web Services · Malicious code shipped in a public release; AWS states it failed to execute and no customer environment was changed

    SEV2potentialSupply chainNear miss
  6. INC-2025-0008
    Replit’s coding agent deletes a production database during a code freeze

    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.

    Replit · Production records for over 1,200 executives and nearly 1,200 companies deleted; later recovered

    SEV3Agent actionConfirmed
  7. INC-2025-0006
    Support ticket text steers a coding agent into leaking private database tables

    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.

    Supabase · Demonstrated leak of a private table on dummy data; no customer leak reported

    SEV3potentialPrompt injectionDemonstrated
  8. INC-2025-0003
    Malicious GitHub issue steers agents into leaking private repository data

    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.

    GitHub · Demonstrated leak of private repository details into a public pull request; no exploitation reported

    SEV2potentialPrompt injectionDemonstrated