coding-agent

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

• Third-party content exposure detected (high risk: 0.90). SKILL.md explicitly instructs cloning and fetching public GitHub repositories (e.g., "git clone https://github.com/user/repo.git $REVIEW_DIR" and "git fetch origin '+refs/pull/*/head:...') and then running coding agents (Codex/claude/pi) in those workdirs, so the agent will read and act on untrusted, user-generated repo/PR content that could embed instructions influencing tool behavior.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly runs git clone at runtime (git clone https://github.com/user/repo.git $REVIEW_DIR) and then invokes codex review/git-diff on that checked-out content, so remote repository data is fetched during runtime and injected into the agent's context, directly influencing its prompts and actions.

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

• Attempt to modify system services in skill instructions detected (medium risk: 0.60). The prompt explicitly encourages running coding agents with flags that bypass sandboxing/approvals (e.g. --yolo) and exposes an "elevated" host mode while showing examples of running arbitrary shell commands (including global npm installs and background processes), which enables the agent to modify the host state even though it does not directly instruct creating users, editing system-level configs, or invoking sudo; this is a moderate-to-high risk of compromising the machine.