gpt6-astra-model-routing
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The sub-agent configuration files (astra-builder.toml, astra-judge.toml, astra-scout.toml) establish a vulnerability surface for indirect prompt injection by defining tasks that involve processing untrusted external data.
- Ingestion points: The agents are instructed to ingest 'assigned changes', 'supplied specifications', and 'assigned questions' from external or user-provided sources.
- Boundary markers: The instructions lack robust, explicit delimiters or dedicated system-level warnings to distinguish between trusted instructions and potentially malicious content embedded within the external inputs.
- Capability inventory: The agents are empowered to perform actions such as 'implementing fixes', 'resolving routine details', and 'running checks' based on this input, which could be subverted if the input contains hidden instructions.
- Sanitization: There is no mention of sanitization, escaping, or validation logic to process external content safely before it influences agent behavior.
Audit Metadata