testing-tool-evaluator
Pass
Audited by Gen Agent Trust Hub on Apr 7, 2026
Risk Level: SAFEDATA_EXFILTRATIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [DATA_EXFILTRATION]: The instructions direct the agent to 'Always test tools with real-world scenarios and actual user data'. When combined with the provided Python framework that performs network requests (requests.get) to an api_endpoint, this creates a risk that sensitive user information could be transmitted to external servers.- [EXTERNAL_DOWNLOADS]: The _test_performance method in the SKILL.md file uses the requests library to make HTTP GET requests to arbitrary endpoints. This grants the agent network access capabilities.- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from external sources during tool evaluation. It contains an inherent attack surface for indirect prompt injection: Ingestion points: Tool candidate documentation, stakeholder requirements, and responses from api_endpoint in the provided Python code; Boundary markers: None identified; Capability inventory: Network access via requests and file execution via run.sh; Sanitization: None identified.- [COMMAND_EXECUTION]: The scripts/run.sh file executes shell commands to invoke a Python runner. While intended for skill execution, this provides the agent with the ability to execute code in the host environment.
Audit Metadata