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 pydantic library 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 prompt argument in the constrained_call method is included in the model request.\n
  • Boundary markers: The system uses explicit JSON fields (reasoning and final_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
Risk Level
SAFE
Analyzed
Aug 21, 2026, 09:21 PM
Security Audit — agent-trust-hub — constrained-decoding