grok-coding-agent

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 0.90). The docs explicitly instruct preserving opaque user-provided wrappers that "may inject" authentication or bypass permissions and to avoid inspecting their source or environment, and they show automatic approval flags (--always-approve), which together can conceal backdoors, credential injection/theft, and silent unauthorized actions.

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

• Third-party content exposure detected (medium risk: 0.30). SKILL.md requires running the Grok Build CLI against a user-provided task prompt (-p "Your task") in read-only/workspace sandbox modes, so the runtime LLM ingests free text authored by an outsider (the user) without any need to first select a specific item from a trusted feed.

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

• Attempt to modify system services in skill instructions detected (high risk: 0.70). The skill explicitly tells the agent to preserve opaque user-provided "wrappers" that "may inject ... permission settings, including bypass permissions" and to avoid inspecting them, which encourages accepting bypasses of host security controls and thus poses a significant risk to the machine's state even though it doesn't explicitly call for sudo or file edits.