soia-dev-task-execute
Pass
Audited by Gen Agent Trust Hub on Aug 6, 2026
Risk Level: SAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from target workspaces and test outputs. It mitigates indirect prompt injection risks by requiring the definition of clear task boundaries and implementing an 'Anti-Fake-Fix Gate' to verify that changes match intended goals. (1) Ingestion points: Target workspace files, test results, and command-line outputs. (2) Boundary markers: Explicit 'Define Task Boundaries' section separating goals from non-goals. (3) Capability inventory: Authorized to modify workspace files and execute verification tests. (4) Sanitization: Privacy guidelines and check gates prevent the inclusion of secrets or irrelevant data in final outputs.
- [EXTERNAL_DOWNLOADS]: The documentation includes commands to install components from the 'soia-team' repositories on GitHub. These are vendor-controlled resources and represent standard delivery methods for the skill.
- [DATA_EXFILTRATION]: Instructions strictly forbid the agent from including credentials, private keys, tokens, or sensitive customer data in task receipts, logs, or persistent storage.
- [COMMAND_EXECUTION]: The skill facilitates the execution of engineering tasks and validation tests, which involves running shell commands within the user's provided workspace environment.
Audit Metadata