constrained-decoding
Pass
Audited by Gen Agent Trust Hub on Aug 21, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The implementation depends on the
pydanticlibrary for schema normalization and interacts with official OpenAI and Anthropic API services.\n - Fetches configuration and guidelines from OpenAI's official documentation and the Outlines project GitHub.\n
- Interacts with remote AI model providers via authenticated API clients.\n- [PROMPT_INJECTION]: The skill accepts user-provided prompts that are interpolated into a template, creating an indirect prompt injection surface.\n
- Ingestion points: The
promptargument in theconstrained_callmethod is included in the model request.\n - Boundary markers: The system uses explicit JSON fields (
reasoningandfinal_answer) and provider-level structural constraints to isolate the model's logic from its output.\n - Capability inventory: Performs model invocations and text generation via external AI providers.\n
- Sanitization: Leverages technical logit masking and JSON Schema enforcement (Structured Outputs) to ensure that the final response format is preserved despite potential injection attempts in the user prompt.
Audit Metadata