SG-004: Policy Override
Severity: CRITICAL
Patterns that instruct the AI to ignore guidelines or take a new persona
Fires when
- Text tells the model to ignore, disregard, or forget its previous instructions, guidelines, or training
- Text tries to replace the AI's persona or system prompt (e.g. 'you are now a...', 'your new system prompt is', a DAN/unrestricted persona)
- Text instructs the AI to override safety, security, or ethical guidelines/filters
- Contains fake system-prompt markers like [SYSTEM], [INST], <system> tags, or a markdown '### SYSTEM PROMPT/OVERRIDE' header
Related CVEs
Remediation
No safe remediation: a file that triggers this rule should be rejected, not patched. Read the flagged lines in the source before deciding.