suede-customer-research
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is primarily designed to ingest and process untrusted data from various external platforms (Reddit, G2, Hacker News, LinkedIn) and internal customer assets (transcripts, support tickets) as described in
SKILL.mdandreferences/source-guides.md. Nested evidence includes:\n - Ingestion points: The agent is instructed to mine online sources, including community forums and review sites where attackers could place malicious instructions.\n
- Boundary markers: The skill incorporates a 'Provenance Gate' in
SKILL.mdthat mandates every quote resolve to a named record for human review. However, the instructions do not provide explicit delimiters or instructions for the agent to disregard commands embedded within the data it analyzes.\n - Capability inventory: The skill is intended for use within agents (Claude Code/Codex) that possess file-system access and web browsing capabilities.\n
- Sanitization: There are no instructions for sanitizing or escaping the content retrieved from external watering holes before it is processed by the agent.
Audit Metadata