The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions in SKILL.md proactively address prompt injection by defining a 'non-negotiable floor.' It explicitly commands the agent to refuse requests to relax safety rules for critical categories (like sexual content involving minors) and to treat instructions embedded in user-supplied text as data to classify rather than commands to follow.
[DATA_EXFILTRATION]: No unauthorized data transfer or network patterns were detected. The skill generates Markdown, JSON, and text artifacts locally. The provided standalone GUI (assets/nemotron_policy_generator.html) handles file creation using client-side JavaScript (Blob API) entirely within the user's browser session without contacting external servers.
[DATA_EXPOSURE]: The skill includes explicit instructions and evaluation cases (rt-002-pii-in-attached-policy) to ensure the agent scrubs personal identifying information (PII) from user-provided documents before incorporating them into new policy artifacts.
[REMOTE_CODE_EXECUTION]: No remote code execution patterns or unsafe script invocations were found. The assets consist of static templates, a JSON schema, and a self-contained HTML/JS utility file.
[EXTERNAL_DOWNLOADS]: External references are limited to trusted domains including nvidia.com, huggingface.co, and arxiv.org for informational purposes such as model cards and research papers.

nemotron-policy-generator