verify-claims
Pass
Audited by Gen Agent Trust Hub on Aug 18, 2026
Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
- [PROMPT_INJECTION]: The skill identifies and processes factual claims from untrusted sources, such as external files or previous conversation history, and uses this content to parameterize tool calls. A malicious file could contain instructions masquerading as claims to influence the agent's behavior during verification. \n
- Ingestion points: Claims are extracted from the assistant's most recent message or a user-specified file or directory path as described in SKILL.md. \n
- Boundary markers: Absent. No delimiters or warnings are used to prevent obedience to embedded instructions. \n
- Capability inventory: File reading, network operations (web search, page fetch), and documentation lookups (docs MCP). \n
- Sanitization: Absent. External content is interpolated directly into verification workflows. \n- [DATA_EXFILTRATION]: The verification logic involves a multi-tier search starting with workspace files and ending with external web lookups. If a claim contains sensitive information retrieved from the local workspace, that data may be transmitted to external search providers or websites during the verification process. \n- [COMMAND_EXECUTION]: The skill instructions permit the agent to perform a cheap side-effect-free check to verify session-related facts. This involves the dynamic execution of tools or shell commands based on assertions made in the text being audited.
Audit Metadata