kf-agent-research-report-only-unless-approved

Pass

Audited by Gen Agent Trust Hub on Mar 30, 2026

Risk Level: SAFE
Full Analysis
  • [PROMPT_INJECTION]: No evidence of malicious prompt injection found. The instructions are designed to restrict agent autonomy and ensure human-in-the-loop confirmation for file system changes.
  • [DATA_EXFILTRATION]: No sensitive file paths or exfiltration patterns detected. The skill permits non-destructive operations like reading files and searching text, which are standard for research tasks.
  • [COMMAND_EXECUTION]: The skill explicitly forbids destructive command execution (like git operations or dependency updates) without prior user consent.
  • [REMOTE_CODE_EXECUTION]: No remote scripts, package installations, or dynamic code execution patterns are present.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes user-provided research requests (ingestion point). While it lacks specific boundary markers for external data, its core logic acts as a defensive measure by defaulting to a 'report-only' mode, significantly reducing the risk of accidental execution of instructions embedded in analyzed files.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 30, 2026, 11:51 AM
Security Audit — agent-trust-hub — kf-agent-research-report-only-unless-approved