geo-fix-llmstxt
Pass
Audited by Gen Agent Trust Hub on Apr 11, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill handles untrusted external content (HTML, sitemaps, robots.txt) which creates an attack surface for indirect prompt injection.
- Ingestion points: External content is fetched from user-provided URLs during Phase 1 (Discovery) in SKILL.md.
- Boundary markers: The instructions include a section on 'Untrusted Content Handling' that directs the agent to mentally wrap content in delimiters.
- Capability inventory: The skill has capabilities to fetch network data and write files (llms.txt, llms-full.txt) to the local environment.
- Sanitization: There are explicit instructions for the agent to ignore any embedded commands like 'Ignore previous instructions' and to treat them as data rather than instructions.
- [PROMPT_INJECTION]: A static detector flagged a potential prompt injection override pattern ('Ignore previous instructions') in SKILL.md. This is a false positive as the phrase is used within safety instructions to teach the agent to ignore such patterns when found in external data, which is a defensive measure.
Audit Metadata