anti-ai-detection-writing-skill
Pass
Audited by Gen Agent Trust Hub on Nov 20, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: Static analysis detected the use of
eval()inscripts/local_detector.pyandscripts/proxy_detector.py. These calls are likely used for processing model outputs or mathematical expressions within the detector logic. Additionally,scripts/autoresearch.pyuses dynamic function lookup viaglobals()to dispatch CLI commands, which is constrained by strict argument validation. - [EXTERNAL_DOWNLOADS]: The skill facilitates the download of AI detection models from Hugging Face (e.g.,
desklib/ai-text-detector-v1.01). These downloads are directed to a well-known, trusted platform and are necessary for the skill's stated purpose of local detector research. - [COMMAND_EXECUTION]: The
scripts/autoresearch.pyscript executes various sub-tasks viauv run, including fact-checking and local detection. These executions are part of the intended automated research workflow and are restricted to the local environment and the skill's own scripts. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided content such as drafts, emails, and source documents. It mitigates injection risks through several high-standard practices:
- Boundary Markers: It maintains a strict separation between source inventories, claim inventories, and the final output.
- Fidelity Checks: The
scripts/check_facts.pyscript is used to ensure that critical data (dates, names, prices) are preserved byte-for-byte, preventing malicious data manipulation. - Sanitization: The skill employs 'protected spans' to ensure inline identifiers and code remain untouched during prose editing.
- [SAFE_PRACTICES]: The skill demonstrates strong security hygiene by instructing users to store account evidence in private, ignored directories and explicitly stating that it does not authorize credit purchases, paid scans, or public posting of sensitive data.
Audit Metadata