submit
Pass
Audited by Gen Agent Trust Hub on May 10, 2026
Risk Level: SAFE
Full Analysis
- [CREDENTIALS_UNSAFE]: The skill follows security best practices for credential management. It references environment variable symbols (handle, email, API tokens) identified in rules/identities.md and explicitly checks if they are set before proceeding, avoiding hardcoded sensitive information.\n- [COMMAND_EXECUTION]: The skill executes a local Python script (brain.py) using uv run to update the internal project database after a successful submission. This is a legitimate part of the skill's workflow and operates on local project files.\n- [DATA_EXFILTRATION]: While the skill is designed to send data to external bug bounty platforms, it includes a mandatory human-in-the-loop safety control. The agent is strictly instructed to show the draft report to the user and obtain explicit confirmation before the submission tool is invoked.\n- [PROMPT_INJECTION]: The skill processes finding details from the local file system (brain/findings/poc). This represents an indirect prompt injection surface where content in findings files could influence the agent. This risk is mitigated by the mandatory user review of the generated draft.\n
- Ingestion points: brain/findings/poc directory, scope.yaml, and findings.json.\n
- Boundary markers: No explicit delimiters are used in the prompt, but the mandatory manual review serves as a functional control boundary.\n
- Capability inventory: File system access (read/write), local script execution, and network submission via MCP tools.\n
- Sanitization: No programmatic sanitization is defined, but the draft generation and user confirmation workflow provides a manual validation layer.
Audit Metadata