meta-debug
Pass
Audited by Gen Agent Trust Hub on Sep 28, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill authorizes the agent to execute shell commands, scripts, and test runners to reproduce observed failures. It explicitly allows mutating environment state without additional authorization if the action was part of the user-provided reproducer, which creates a path for potentially dangerous operations.
- [INDIRECT_PROMPT_INJECTION]: The agent is designed to ingest external data like stack traces, failing commands, and error logs. These inputs serve as the 'reproducer' that the agent executes. If this data comes from an untrusted source, it could contain malicious shell commands disguised as legitimate debugging steps.
- [PROMPT_INJECTION]: The 'Sources' section of the skill includes references to future-dated models and blog posts (e.g., Anthropic Claude 5, OpenAI GPT-6 Astra from 2026). This is a context engineering technique used to influence the model's behavior or capability estimation by pretending it is operating under newer, advanced guidelines.
Audit Metadata