pr-review-canvas

Pass

Audited by Gen Agent Trust Hub on Aug 19, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill invokes shell commands such as gh api and python3 using variables like owner, repo, and PR number extracted from user input. While standard for this use case, it constitutes a potential injection point if the agent fails to validate these inputs correctly.
  • [PROMPT_INJECTION]: The skill exhibits a vulnerability surface for indirect prompt injection (Cross-Site Scripting). It processes untrusted data from GitHub PRs (descriptions and comments) and embeds them into a generated HTML file.
  • Ingestion points: Untrusted data enters via gh api calls fetching PR details and comments (SKILL.md, Step 1).
  • Boundary markers: Although the skill instructions include techniques for safe JSON injection, they do not prescribe boundary markers or mandatory escaping for the free-form HTML body generated by the agent.
  • Capability inventory: The skill can write files to the local filesystem and start a local web server to display them (SKILL.md, Step 4).
  • Sanitization: A basic HTML escaping function is provided in renderer.js for diff rendering, but its application to other PR-derived content in the agent-written HTML is not explicitly required, allowing potentially malicious scripts from a PR to execute in the user's browser.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 19, 2026, 03:44 PM
Security Audit — agent-trust-hub — pr-review-canvas