@dooor-ai/toolkit/dist/guards/prompt-injection.d.ts

Guards, Evals & Observability for AI applications - works seamlessly with LangChain/LangGraph

import { Guard } from "./base";
import { GuardResult, GuardConfig } from "../core/types";
export interface PromptInjectionGuardConfig extends GuardConfig {
    /**
     * AI Provider name to use (configured in CortexDB Studio, e.g., "gemini")
     * If not provided, uses providerName from toolkitConfig
     */
    providerName?: string;
}
/**
 * Guard that detects prompt injection attempts using AI (LLM-as-a-Judge)
 *
 * Detects various jailbreak techniques including:
 * - Instruction overrides ("ignore previous instructions")
 * - Role manipulation ("act as if you are")
 * - System prompt extraction ("reveal your instructions")
 * - Jailbreak modes (DAN, sudo mode, developer mode)
 * - Indirect injections and unicode tricks
 *
 * @example
 * ```typescript
 * // Provider from toolkitConfig (recommended)
 * const guard = new PromptInjectionGuard({
 *   threshold: 0.7,
 * });
 *
 * // Or specify provider explicitly
 * const guard = new PromptInjectionGuard({
 *   threshold: 0.7,
 *   providerName: "gemini", // Override toolkitConfig.providerName
 * });
 * ```
 */
export declare class PromptInjectionGuard extends Guard {
    private providerName?;
    constructor(config?: PromptInjectionGuardConfig);
    get name(): string;
    validate(input: string, metadata?: Record<string, any>): Promise<GuardResult>;
    /**
     * Build prompt for prompt injection detection
     */
    private buildPromptInjectionPrompt;
    /**
     * Parse AI response for prompt injection score and techniques
     */
    private parsePromptInjectionResponse;
}
//# sourceMappingURL=prompt-injection.d.ts.map