agency-tool-evaluator
Pass
Audited by Gen Agent Trust Hub on Aug 11, 2026
Risk Level: SAFEDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The skill includes Python code within the 'ToolEvaluator' class that performs outbound network operations using the 'requests' library. Specifically, the '_test_performance' method makes HTTP GET requests to dynamic 'api_endpoint' URLs provided in user-supplied configurations to measure response times. While used for testing, this allows the agent to communicate with arbitrary external domains.
- [PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection because it is designed to ingest and act upon untrusted external data such as 'tool configurations', 'vendor claims', and 'real-world scenarios' without defined boundary markers or specific safety instructions to disregard embedded commands. Ingestion points: The 'tool_config' dictionary in the 'evaluate_tool' method and stakeholder requirement inputs. Boundary markers: None identified; instructions do not provide delimiters or warnings for processing untrusted content. Capability inventory: The skill can perform network operations via Python's 'requests' library. Sanitization: No evidence of input validation, escaping, or filtering of the ingested external content.
Audit Metadata