negative-keywords

Pass

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted search term data and external website content, which presents a surface for indirect prompt injection attacks.\n
  • Ingestion points: Untrusted data enters the agent context through search-term-report outputs (CSV or pasted lists) and external website content retrieved during the research phase mentioned in SKILL.md.\n
  • Boundary markers: The skill implements strong boundary instructions, explicitly stating that untrusted content must be treated as data rather than instructions. It directs the agent to quote any detected instructions within data rather than following them (references/agent-security.md).\n
  • Capability inventory: Across SKILL.md and referenced files, the agent is permitted to fetch external URLs and write research findings to the local .agents/product-context.md file.\n
  • Sanitization: The skill mandates HTML-escaping of all interpolated values and strict validation of URLs to use https: or relative paths before they are written into any markup or code output (references/agent-security.md).\n- [DATA_EXPOSURE]: The skill includes robust guidelines to prevent the exposure or exfiltration of sensitive information.\n
  • The instructions explicitly forbid writing credentials (API keys, tokens, passwords) into files, deliverables, or commits, and require redaction of tokens from URLs or input data (references/agent-security.md).\n- [PROMPT_INJECTION]: The skill contains instructions that prioritize clean, upload-ready output by directing the agent to stop and ask for missing information rather than producing a draft with warnings (SKILL.md).\n
  • This instruction is designed to prevent user error during the copy-paste process into advertising platforms and does not bypass safety filters or conceal malicious activity.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 17, 2026, 02:23 PM