safety-scan

Pass the untrusted text to the AIDefence is_safe tool for a quick boolean verdict, or use the analyze tool for detailed threat classification. Run this check on user submissions, API payloads, and webhook data before model processing.

The scan detects direct and indirect prompt injection, jailbreak attempts, data exfiltration patterns, instruction override attacks, and social engineering prompts. Each threat is classified with confidence scores during deep analysis.

Yes, confirmed threats can be submitted to the AIDefence learn tool to improve future detection. You can also review detection rates and false positive metrics through the stats tool.

Use the quick is_safe check for high-throughput screening where a boolean verdict suffices. Use the full multi-layer scan or deep analysis when handling high-risk inputs or when you need detailed threat classification and confidence scores.

Automated scanning can produce false positives on legitimate but unusual inputs and may miss novel attack patterns not yet learned. Combine scanning with least-privilege tool permissions and human review for high-stakes actions.