The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user tasks and model-generated answers by passing them to an external critic model. Ingestion points include the task, answer, and context fields in the generated JSON packets. The critic instructions in scripts/duel.py explicitly warn the external model that packet fields are untrusted data and should not be treated as instructions. The duel.py script has network access to a well-known API and file-writing capabilities within a dedicated run directory. Input packets are validated for size and structure by the validate_packet function to ensure integrity.\n- [COMMAND_EXECUTION]: The skill executes a bundled Python script (duel.py) to handle API interactions and file management. Commands are constructed with explicit flags and avoid raw interpolation of user arguments into shell strings, preventing command injection vulnerabilities.\n- [EXTERNAL_DOWNLOADS]: The skill communicates with OpenAI's official API to facilitate the debate. The duel.py script sends data to a well-known service endpoint (https://api.openai.com/v1/responses) and incorporates security best practices such as disabling HTTP redirects and enforcing strict JSON parsing to prevent credential leakage.

duel