adversarial-verify

Pass

Audited by Gen Agent Trust Hub on Aug 11, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its ingestion of untrusted findings and reproduction commands.
  • Ingestion points: Content enters the agent context through the $claim argument and findings list paths (e.g., in SKILL.md Step 1), which are then passed to sub-agents as ground truth evidence.
  • Boundary markers: The verifier prompt template in references/verify-patterns.md uses simple text headers but lacks secure delimiters or explicit instructions to treat data as non-executable, increasing the risk of instruction confusion.
  • Capability inventory: The skill grants verifier agents access to the Bash tool and explicitly instructs them to execute reproduction commands provided in the untrusted data (SKILL.md Step 3).
  • Sanitization: No input validation, filtering, or escaping is performed on the content of the claims before they are processed by agents or executed as commands.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 11, 2026, 06:20 PM
Security Audit — agent-trust-hub — adversarial-verify