meta-skill-architect
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The evaluation suite in
scripts/test_runner.pyusessubprocess.runto call theclaudeCLI tool for grading skill outputs. This execution is an intended part of the development and testing workflow. - [DYNAMIC_EXECUTION]: The skill generates new Python (.py) and Shell (.sh) scripts dynamically based on the identified needs of the skill being designed, as specified in the S1 and S3 protocols in
references/protocols-advanced.md. - [INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests untrusted user prompts to generate instruction modules. It incorporates defensive measures to mitigate this risk.
- Ingestion points: User input for skill creation (task.md) and informal prompt migration (references/protocols-advanced.md).
- Boundary markers: The system requires the use of semantic delimiters like
<input>,<data>, and<ticket>to separate data from instructions. - Capability inventory: File writing and script generation.
- Sanitization: Identity protection rules in
system.mdexplicitly instruct the agent to halt and report if an input attempts to override its core safety guidelines or identity.
Audit Metadata