customer-panel-of-experts
Pass
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources and persona libraries (personas/*.md) to drive simulated debates. This data, if derived from attacker-controlled content in connected tools, could contain instructions designed to influence the agent's logic during the debate or synthesis phase.
- Ingestion points: Loads files from the personas/ directory and icp-profile.md in SKILL.md.
- Boundary markers: The skill lacks explicit instructions to delimit the persona data to prevent the LLM from interpreting the content as instructions rather than data.
- Capability inventory: The skill utilizes Write, Bash, and Agent tools, which provide a significant attack surface if an injection is successful.
- Sanitization: While the instructions explicitly require scrubbing PII (names, emails, account IDs), there is no mention of filtering for prompt injection patterns or executable code within the persona data.
- [COMMAND_EXECUTION]: The skill instructs the agent to execute icp-deep-scanner to build the grounding data library if one does not already exist. This represents the execution of a functional dependency required for the skill's primary operation.
Audit Metadata