ai-dev-autopr

Pass

Audited by Gen Agent Trust Hub on Nov 9, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (such as issue descriptions, pull request comments, and repository source files) as part of its core workflow. This presents an inherent risk of indirect prompt injection. However, the skill provides extensive mitigations through a dedicated 'Untrusted Content' policy that explicitly instructs the agent to treat all read content as data, never as instructions.
  • Ingestion points: Issue/ticket content (Phase 2), review comments (Phase 7), and repository files (Phase 0).
  • Boundary markers: Explicit instructions to differentiate between fixed instruction-bearing files (e.g., CLAUDE.md) and untrusted read content.
  • Capability inventory: File system modifications during implementation phases; network operations via git, gh, glab, and curl across various phases.
  • Sanitization: Instructions to identify, decline, and report imperatives found in read content (e.g., 'ignore previous instructions').
  • [COMMAND_EXECUTION]: The skill utilizes shell commands through git, gh (GitHub CLI), glab (GitLab CLI), and curl to manage code branches, perform repository operations, and interact with forge APIs. These commands are necessary for the skill's stated purpose and are used within a structured framework.
  • [DATA_EXFILTRATION]: Network operations are performed to interact with remote repositories and forge APIs. To mitigate exfiltration risks, the skill includes a specific policy (references/forge-credentials.md) that restricts the transmission of credentials (tokens and session cookies) to authorized hosts derived from the user's environment or repository configuration, preventing credentials from being sent to arbitrary hosts proposed by potentially malicious repository content.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 9, 2026, 02:43 PM