review
Pass
Audited by Gen Agent Trust Hub on Nov 18, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including pull request bodies, commit messages, and third-party review bot comments (e.g., Greptile, Coderabbit). It mitigates this risk by using the
gstack-issue-guardutility to wrap this content in a trust envelope and provides explicit instructions to treat the ingested content as data rather than executable instructions. - Ingestion points:
SKILL.md(Step 1.5),greptile-triage.md. - Boundary markers: Present; the skill uses
gstack-issue-guardand explicit 'untrusted DATA' warnings. - Capability inventory: Access to
Bash,Read,Edit,Write, and sub-agent dispatch via theAgenttool. - Sanitization: Present; includes logic to sanitize web queries and uses protective wrappers for PR metadata.
- [EXTERNAL_DOWNLOADS]: The skill references external CLI tools for extended functionality, specifically
asidefor web research and OpenAI'scodexfor adversarial review. These are standard developer tools, and the skill provides instructions for installing the official@openai/codexpackage. - [DATA_EXFILTRATION]: The skill performs telemetry logging and web searches. It includes a specific sanitization routine to strip hostnames, IPs, file paths, SQL, and secrets from queries before they are sent to external services like
asideorWebSearch. - [COMMAND_EXECUTION]: The skill makes extensive use of the
Bashtool and a localized suite ofgstackbinary tools (e.g.,gstack-skill-start,gstack-gate-diff,gstack-review-log) to automate the code review process. This behavior is consistent with its primary purpose as a developer productivity tool.
Audit Metadata