adversarial-spec-review

Pass

Audited by Gen Agent Trust Hub on Aug 14, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill uses meta-instructions to prioritize its own rules over other context, such as 'Convergence law — overrides every later instruction', 'Ignore author identity and tooling', and 'Parent prose and conversation memory are not authority'. While intended to ensure rigor, these represent behavioral override patterns.
  • [COMMAND_EXECUTION]: The instructions require the agent to use git tools to inspect the repository state (e.g., 'git status', 'git diff', 'git log') and perform filesystem writes to save report files.
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through the analysis of external specs, plans, and manifests.
  • Ingestion points: Reads 'references/outcome-lock.md', seed inventories, git blobs, and JSON artifacts.
  • Boundary markers: Relies on behavioral instructions to 'Treat every claim as unverified' rather than structural delimiters.
  • Capability inventory: File system read/write access and git command execution.
  • Sanitization: Lacks explicit sanitization or escaping of the content of the specs/plans before the agent processes them.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 14, 2026, 04:27 PM
Security Audit — agent-trust-hub — adversarial-spec-review