diagnosing-bugs
Pass
Audited by Gen Agent Trust Hub on Oct 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as network traces, log dumps, and HAR files to diagnose bugs. This provides an attack surface where malicious content within these artifacts could attempt to influence the agent's behavior. \n
- Ingestion points: Captured traces, HAR files, log dumps, and core dumps (SKILL.md).\n
- Boundary markers: Absent.\n
- Capability inventory: Shell execution (bash), network operations (curl), browser automation (Playwright/Puppeteer), and test runners (SKILL.md).\n
- Sanitization: Explicit requirement to redact secrets and credentials using (SKILL.md).\n- [DYNAMIC_EXECUTION]: The instructions direct the agent to generate and execute various scripts at runtime, including test harnesses, replay scripts, and bisection loops. It also utilizes a template for human-in-the-loop scripts (scripts/hitl-loop.template.sh).\n- [DATA_EXFILTRATION]: The skill includes instructions to use curl for constructing feedback loops against development servers. It proactively addresses potential data exposure by requiring the agent to redact all secrets and use environment variables for credentials rather than hardcoding them in loops or captured artifacts.
Audit Metadata