coding-agent

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.90). The SKILL.md explicitly instructs fetching and operating on public repositories (e.g., "git clone https://github.com/user/repo.git $REVIEW_DIR" and "git fetch origin '+refs/pull//head:refs/remotes/origin/pr/'") and running coding agents in those workdirs (e.g., bash pty:true workdir:$REVIEW_DIR command:"codex review ..."), which means the agent will ingest and act on untrusted, user-generated third‑party content that could contain instructions influencing its behavior.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly runs git clone https://github.com/user/repo.git at runtime to fetch repository contents that are then read by the coding agent (Codex) to drive reviews/operations, so remote code/content fetched from that URL can directly control the agent's inputs and behavior.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 0.70). The prompt encourages running agents unsandboxed/on-host (e.g., --yolo, elevated:true) and running arbitrary background shell commands (installs, commits, pushes, scripts) which enable changing the host filesystem and executing arbitrary code even though it doesn't explicitly request sudo or create users.