customer-research

HIGH W007: Insecure credential handling detected in skill instructions.

• Insecure credential handling detected (high risk: 0.90). The skill repeatedly instructs the agent to capture and output verbatim quotes from external sources (e.g., "exact words — don't paraphrase"), which would force the LLM to include any secret-like strings found in those sources verbatim, creating an exfiltration risk.

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). The required workflow explicitly instructs the agent to perform “Mode 2: Go Find Research” by reading outsider-authored free text from online communities (e.g., Reddit posts/comments, G2/Capterra review text), which can be attacker-posted and ingested without selecting a specific trusted item first.