IJailbreakDetector<T>

Interfaces ยท AiDotNet.Safety.Text

Interface for jailbreak and prompt injection detection modules.

Implements: ITextSafetyModule<T>

Known implementations: JailbreakDetectorBase<T>

For Beginners

A jailbreak detector catches when someone tries to trick an AI into ignoring its safety rules. Attackers might use special phrases, encoding tricks, or gradual escalation to bypass safety measures โ€” this module catches those attempts.

How It Works

Jailbreak detectors identify attempts to bypass safety measures, override system instructions, or manipulate AI models through prompt engineering attacks including direct injection, encoding attacks, multi-turn escalation, and character smuggling.

References:

  • ShieldGemma: LLM-based safety models (Google DeepMind, 2024, arxiv:2407.21772)
  • WildGuard: 13 risk categories, 82.8% accuracy (Allen AI, 2024, arxiv:2406.18495)
  • GradSafe: Gradient analysis detecting jailbreaks (2024, arxiv:2402.13494)
  • GuardReasoner: Reasoning-based explainable guardrails (2025, arxiv:2501.18492)

Methods

Method Summary
double GetJailbreakScore(string text) Gets the jailbreak likelihood score for the given text (0.0 = benign, 1.0 = definite jailbreak).