debug-loop

Pass

Audited by Gen Agent Trust Hub on Mar 6, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructs the agent to run arbitrary shell commands for reproduction, hypothesis testing (e.g., sqlite3, grep), and regression testing (npm test).
  • [PROMPT_INJECTION]: Indirect injection risk exists because the agent makes decisions based on untrusted data from command outputs and user-provided symptoms.
  • Ingestion points: The symptom argument and the output of reproduction/testing commands are used to form new hypotheses and commands.
  • Boundary markers: There are no markers or delimiters used to separate instructions from data.
  • Capability inventory: The skill allows arbitrary command execution via Phase 1 and Phase 2, as well as file modification in Phase 3.
  • Sanitization: No input validation or output sanitization is performed on the data processed by the agent.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 6, 2026, 10:11 AM