sahit-sai prompt-fence
Purpose
Catch risky prompt text before it gets bundled into skills, system prompts, or operator playbooks.
Runbook
- Scan one file, one directory, or a narrowed skill subtree rather than an entire home directory.
- Run
scripts/scan.pyto flag common jailbreak phrases, data exfiltration requests, and unsafe shell habits. - Review each finding in context; a match is a lead for review, not automatic proof of malicious intent.
- Escalate only the lines that clearly weaken instruction boundaries or leak secrets.
Stop conditions
- Abort if the scan target includes private data that should not be copied into reports.
- Abort before classifying a file as malicious without manual review.
- Abort if the current variant cannot safely inspect the relevant local files.
Output format
- File-by-file finding inventory
- Severity, rule name, line number, and matched excerpt
- A concise markdown or JSON review artifact
Example invocations
python3 skills/prompt-fence/scripts/scan.py skills/python3 skills/prompt-fence/scripts/scan.py README.md --markdown