empirical-prompt-tuning
Pass
Audited by Gen Agent Trust Hub on Aug 21, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill's workflow involves dispatching subagents to execute and evaluate user-provided prompts. This exposes the subagent to indirect prompt injection risks if the prompt being tuned contains malicious instructions designed to hijack the subagent's execution context.\n
- Ingestion points: The 'target prompt' and 'scenario' content are interpolated into the subagent's system instructions in the 'Subagent invocation contract' section of SKILL.md.\n
- Boundary markers: While the contract uses Markdown headers for structure, it lacks explicit isolation markers or security directives to prevent the subagent from obeying instructions embedded within the target prompt.\n
- Capability inventory: The subagent is dispatched using the 'Task' tool, granting it potential access to the agent's full toolset, including file system and shell operations.\n
- Sanitization: The skill does not define specific sanitization or escaping logic for the untrusted prompt content before it is passed to the subagent.
Audit Metadata