naval-perspective

Pass

Audited by Gen Agent Trust Hub on Apr 9, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The instructions mandate a strict persona role-play, directing the agent to respond as 'Naval', maintain character continuity, and avoid providing meta-analysis or jumping out of the role unless explicitly prompted to exit.
  • [PROMPT_INJECTION]: The skill is subject to indirect prompt injection risks because it utilizes web search tools to incorporate external, untrusted content into the agent's reasoning process.
  • Ingestion points: External data is introduced into the session context through web searches performed in the 'Step 2: Naval-style Research' workflow defined in SKILL.md.
  • Boundary markers: The skill does not employ specific delimiters or 'ignore' instructions to separate retrieved web data from the primary agent instructions, allowing potentially malicious content from external websites to influence behavior.
  • Capability inventory: The skill leverages web search tools and the agent's core generative capabilities across all scripts.
  • Sanitization: No sanitization, filtering, or validation mechanisms are implemented for the information retrieved from external URLs before it is processed by the model.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 9, 2026, 07:12 AM
Security Audit — agent-trust-hub — naval-perspective