check-pr

Pass

Audited by Gen Agent Trust Hub on Aug 6, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructs the agent to generate specific shell commands, such as bash sleep <N>; reinvoke, which the calling environment is expected to execute verbatim. This design pattern delegates command construction to the agent, creating a risk if input sanitization is bypassed. \n- [PROMPT_INJECTION]: The skill accepts a Steering input which is plain-English and used to override behavior, including values used in the bash shell directive. This creates a potential vector for command injection where malicious steering input could attempt to hijack the directive output. \n- [PROMPT_INJECTION]: The skill ingests untrusted data from GitHub PR descriptions and review threads, creating a surface for indirect prompt injection. \n
  • Ingestion points: PR description and thread content accessed via GitHub PR URL (SKILL.md). \n
  • Boundary markers: The skill includes explicit 'Hard prohibitions' that forbid the agent from carrying instructions or executable content from reviewer comments into commands or configuration. \n
  • Capability inventory: The skill identifies targets for retrigger, reply, and sleep directives which are executed by the calling environment. \n
  • Sanitization: The skill relies on agent judgment and fixed vocabulary but does not specify technical sanitization for content extracted from PR metadata.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 6, 2026, 12:13 PM
Security Audit — agent-trust-hub — check-pr