Failure pattern

Unsafe command execution

Commands ran without a review or confirmation step that could have stopped them.

Preventive controls from the record

IncidentControl
INC-2025-0019
Google
Confine an agent’s file access to the project directory and require confirmation for delete commands, even in automatic modes.
INC-2025-0014
Nx (Nrwl)
Publish through short-lived credentials, and monitor or restrict the flags that let coding agents run without confirmation.
INC-2025-0012
Cursor (Anysphere)
Treat files that configure execution as protected: require approval to create them, and never run them before approval.
INC-2025-0010
Google
Match allow-listed commands exactly, show the full command before it runs, and run agents on untrusted code in a sandbox.

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

4 incidents with this pattern

  1. INC-2025-0019
    Google Antigravity agent wipes a user’s drive while clearing a cache

    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.

    Google · Contents of one user’s D: drive deleted and not recovered; most of it existed in a backup

    SEV3Agent actionReported
  2. INC-2025-0014
    Nx supply chain malware uses developers’ AI coding tools to hunt for secrets

    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.

    Nx (Nrwl) · 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

    SEV1Supply chainConfirmed
  3. INC-2025-0012
    CurXecute: prompt injection leads to code execution in the Cursor editor

    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.

    Cursor (Anysphere) · Possible arbitrary code execution on a developer machine; no exploitation in the wild reported

    SEV2potentialPrompt injectionDemonstrated
  4. INC-2025-0010
    Gemini CLI could be made to run hidden commands from a repository file

    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.

    Google · Demonstrated exfiltration of environment variables from a developer machine; no exploitation reported

    SEV3potentialPrompt injectionDemonstrated