Threat Defense

Threat Defense provides real-time protection against malicious prompt attacks, jailbreaks, instruction leaks, and output manipulation.

Protected Threat Vectors

Threat	Description
Prompt Injection	Attempts to override or bypass system instructions and constraints.
Jailbreaks	Adversarial prompts designed to circumvent safety guardrails.
Instruction / System Prompt Leakage	Extraction attacks attempting to reveal proprietary instructions.
Role Drift / Hijacking	Manipulating the assistant to deviate from its intended business domain.
Data Extraction Attacks	Automated attempts to harvest confidential training or index data.
Canary Token Leaks	Detection of hidden canary strings signaling instruction exposure.

Evaluates every incoming prompt before it reaches the language model:

Allow: Request passes all safety checks.
Flag: Request is permitted, but logged as a security event for administrative review.
Block: Request is blocked immediately with a signed denial event.

Monitors model outputs prior to client delivery to catch:

Canary tokens are unique, hidden cryptographic markers inserted into chatbot instructions:

How it works: If an attacker succeeds in extracting instructions, the canary marker in the response triggers an immediate high-priority security alert.
Configuration: Go to Admin ? Threat Defense ? Canary Tokens to generate, rotate, or monitor tokens.

Configure Policies: Go to Admin ? Threat Defense ? Policies. Select from Strict (public chatbots), Balanced (standard enterprise use), or Permissive (testing environments).
Review Incidents: Inspect flagged or blocked events under Admin ? Threat Defense ? Incidents to analyze attack payloads and adjust sensitivity.