ai-security

To detect prompt injection and jailbreak attempts, run a Python-based threat scanner that statically analyzes LLM prompts, tool calls, and retrieval-augmented contexts for malicious patterns and data-extraction signals.

MITRE ATLAS mapping works by linking detected prompt injection and jailbreak patterns to specific adversary techniques like AML.T0051, AML.T0056, and AML.T0020, providing structured risk scoring and guardrail design recommendations.

Yes, you can scan retrieval-augmented contexts for data exfiltration risks. The threat scanner evaluates prompts and tool calls across these interactions to identify data-extraction signals before deployment.

The best way to assess AI assistant exposure is to run a threat scanner on your seed prompt dataset, generating JSON output with risk scores and actionable guardrail recommendations to harden system security.

No external dependencies are required to run static pattern scanning for AI threats. The skill operates using a standalone Python-based threat scanner with built-in authorization checks and JSON output generation.

Automated risk assessment workflows for jailbreak detection rely on static pattern scanning, meaning they may miss novel or obfuscated injection vectors that do not match pre-defined MITRE ATLAS techniques or known data-extraction signals.