researching-competitor-ads

Pass

Audited by Gen Agent Trust Hub on Sep 1, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external advertisements which could theoretically contain malicious instructions.
  • Ingestion points: Untrusted ad data is fetched via the social_research tool in SKILL.md (Step 3) and references/pulling-competitor-ads.md.
  • Boundary markers: Vision model prompts in references/reading-one-ad.md use a fixed narrative framework to structure the analysis, reducing the likelihood of the model obeying instructions within the ad content.
  • Capability inventory: The agent executes a local script build_ledger.py on the fetched data and uses vision tools (video_analysis, image_analysis) to interpret media content.
  • Sanitization: The build_ledger.py script uses standard library JSON parsing and simple string normalization.
  • [COMMAND_EXECUTION]: The skill utilizes a local Python script to aggregate and process research data.
  • Evidence: SKILL.md Step 4 executes python3 <skill-dir>/scripts/build_ledger.py using arguments derived from previous tool outputs.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 1, 2026, 11:10 AM
Security Audit — agent-trust-hub — researching-competitor-ads