benchmark-due-diligence
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data fetched from the internet via
WebSearch,WebFetch, and specialized CLI tools. This creates an attack surface where malicious instructions embedded in web pages could influence agent behavior. - Ingestion points: Data is gathered from the open web and social platforms (
agent-reach,qcc) in the initial collection phase. - Boundary markers: The skill employs a strict "two-channel split" architecture, separating public
FACTSfrom privateCOMMISSIONER_CONTEXT. It also uses JSON schemas to structure agent outputs. - Capability inventory: The skill possesses the ability to execute shell commands (via the
Workflowtool), perform network requests, and manage a multi-agent orchestration pipeline. - Sanitization: While the skill uses JSON schemas to ensure structured output, it does not explicitly detail sanitization or escaping of the raw HTML/text content retrieved from the web before it is analyzed by the agents.
- [COMMAND_EXECUTION]: The orchestration template provided in the skill references the use of
Bashto invoke external tools likeagent-reachfor social media data andqccfor corporate records. This represents an intended use of local command execution for data enrichment.
Audit Metadata