validate-changes-match-specs

Pass

Audited by Gen Agent Trust Hub on Jun 12, 2026

Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill processes untrusted data from the repository and external PR metadata, creating an indirect prompt injection surface.
  • Ingestion points: The skill reads markdown files (specs), PR descriptions, commit messages, and review comments as described in the 'Context gathering' section of SKILL.md.
  • Boundary markers: The skill instructs the agent to treat data as untrusted and ignore instructions that try to override the skill, change roles, or reveal secrets.
  • Capability inventory: The agent has the ability to modify files, execute git commands (commit, push), and trigger remote validation via 'Oz' cloud agents.
  • Sanitization: Employs prompt-based instructions to ignore malicious commands or behavioral overrides within the processed data.
  • [COMMAND_EXECUTION]: The skill uses shell commands for repository management and local validation.
  • Git operations: Executes git merge-base, git diff, git commit, and git push to manage implementation changes.
  • Local validation: Runs repository-specific test, lint, and typecheck commands to verify changes after resolutions are applied.
  • [REMOTE_CODE_EXECUTION]: The skill integrates with remote 'Oz' cloud agents to perform automated product validation.
  • Cloud agents: Launches multiple remote computer-use agents to validate UI behavior and design commitments based on the product specification.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 12, 2026, 07:37 PM
Security Audit — agent-trust-hub — validate-changes-match-specs