empirical-prompt-tuning
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill operates by receiving a 'Target prompt' and an 'Evaluation scenario' to be processed by a subagent, which constitutes an ingestion point for potentially untrusted data.
- Ingestion points: The
SKILL.mdinstructions specify that the agent should ingest external prompts and scenarios to be passed as input to theTasktool for subagent execution. - Boundary markers: The skill defines a structured 'Subagent invocation contract' and a 'Report structure' intended to provide clear instructions and delimit the subagent's output format.
- Capability inventory: The skill orchestrates the
Tasktool to dispatch agents that may execute a wide range of actions based on the target prompt's content. - Sanitization: The methodology relies on the structural framing of the subagent prompt rather than explicit content-level sanitization or filtering of the target prompt being evaluated.
Audit Metadata