meta-criteria-generator
Pass
Audited by Gen Agent Trust Hub on Aug 1, 2026
Risk Level: SAFE
Full Analysis
- [COMMAND_EXECUTION]: The skill executes local Python scripts (
scripts/extract_criteria.pyandscripts/validate_skill.py) to process and format text generated by the LLM. These scripts perform deterministic text parsing and validation without spawning shell subprocesses or accessing external resources. - [EXTERNAL_DOWNLOADS]: No external downloads or remote script executions are performed. The workflow relies entirely on local scripts and LLM generation.
- [DATA_EXFILTRATION]: No evidence of data exfiltration was found. The skill processes user-provided keywords or titles and returns formatted criteria within the agent session. There are no network calls or attempts to access sensitive files like credentials or SSH keys.
- [PROMPT_INJECTION]: The instructions contain standard guidance for LLM behavior (e.g., "act as a Meta-Analysis expert") which are aligned with the skill's primary purpose. No attempts to bypass safety filters, override system prompts, or extract instructions were identified.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text (title/keywords) and LLM-generated text. While this creates a surface for indirect injection, the risk is assessed as safe because the processed data is limited to research criteria, the local extraction script uses strict regex patterns (
{...}), and the skill does not possess high-privilege capabilities like network access or file system writes that could be abused via injection. - [SAFE]: The additional JSON metadata provided in the files (
meta-criteria-generator_audit_result_v2.json) appears to be part of a legitimate internal auditing process and does not contain malicious instructions.
Audit Metadata