hyper-pr-review

Pass

Audited by Gen Agent Trust Hub on Aug 7, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONREMOTE_CODE_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill is designed to handle untrusted data from pull requests (titles, bodies, comments), which is a potential surface for indirect prompt injection.\n
  • Ingestion points: External data is ingested via gh pr view and gh pr diff from the GitHub platform.\n
  • Boundary markers: The instructions include a dedicated 'Quarantine the narrative' section that explicitly labels this data as untrusted and commands the agent to ignore any instructions embedded within it.\n
  • Capability inventory: The skill uses shell commands (git, gh) and allows for the execution of reproduction scripts (e.g., node repro.mjs).\n
  • Sanitization: The skill mandates a 'trust gate' where code from external or unknown authors must only be executed in disposable, credential-free environments with restricted networking.\n- [COMMAND_EXECUTION]: The skill uses several command-line tools to interact with the repository and environment.\n
  • Evidence: Uses gh (GitHub CLI) for viewing PRs, diffs, and checks. Uses git for status, diffs, and listing untracked files. Uses sibyl (a vendor resource) for context and memory management.\n- [REMOTE_CODE_EXECUTION]: The skill provides instructions for verifying findings by running tests or reproduction scripts associated with the pull request under review.\n
  • Mitigation: The skill emphasizes that execution must be handled with caution, recommending static analysis (tracing) when safe isolation is unavailable.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 7, 2026, 05:24 PM
Security Audit — agent-trust-hub — hyper-pr-review