skill-creator
Pass
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses the
subprocessmodule inscripts/run_eval.py,scripts/improve_description.py, andeval-viewer/generate_review.pyto execute system commands. Specifically, it calls theclaudeCLI tool to run evaluation queries and optimize descriptions, and thelsofutility to manage local network ports for its evaluation viewer. - [DYNAMIC_EXECUTION]: In
scripts/run_eval.py, the skill dynamically generates and writes markdown command files to the.claude/commands/directory. These files are used to inject specific skill metadata and descriptions into the execution context of theclaude -pCLI tool during evaluation runs. - [INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves ingesting and executing user-provided test prompts and input files through subagents. This creates a vulnerability surface for indirect prompt injection.
- Ingestion points: Untrusted data enters the context via user-defined prompts in
evals/evals.jsonand associated input files. - Boundary markers: The instructions do not define strict delimiters or specific 'ignore embedded instructions' warnings for these external prompts.
- Capability inventory: The skill can execute shell commands via
subprocess, write to the filesystem, and spawn new subagent tasks. - Sanitization: There is no evidence of sanitization or filtering for the content of user-provided test prompts before they are interpolation into the subagent's task description.
Audit Metadata