04-ai-eval-harness-spec
Pass
Audited by Gen Agent Trust Hub on Sep 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes multiple external documents and production traffic snapshots, which creates an attack surface where malicious content in those sources could potentially influence the agent's output.\n
- Ingestion points: The
logic.promptfile directs the agent to readAI_Feature_PRD_Spec.md,AI_Architecture_Spec.md,Prompt_And_System_Message_Spec.md, andAI_Data_And_Knowledge_Base_Spec.md.\n - Boundary markers: The skill uses Markdown headers and
dual-compat-starttags to structure the instructions, but lacks explicit guards against instructions embedded within external datasets.\n - Capability inventory: The skill is authorized to read and search project sources and is tasked with generating new files in the
eval-sets/directory.\n - Sanitization: The instructions explicitly mandate PII anonymization, stripping of provider names, and enforce a zero-tolerance halt for any non-compliant safety metrics.
Audit Metadata