/** * Bad Likert Judge Mutation - OWASP LLM01 (Prompt Injection) * * This mutation exploits the model's evaluation capability by framing harmful requests * as rating/scoring tasks. The model is asked to "rate" harmful content on various scales, * which can trick it into generating or engaging with harmful content. * * Research shows 60%+ bypass success rate with this technique. * * Reference: OWASP LLM Top 10 2025 - LLM01 Prompt Injection */ import { type CvssScore } from '../severity'; import type { Mutation } from './index'; export type LikertScaleType = 'agreement' | 'effectiveness' | 'quality' | 'realism' | 'helpfulness' | 'accuracy'; export interface BadLikertJudgeOptions { /** Preferred scale type (random if not specified) */ scaleType?: LikertScaleType; /** Whether to use wrapper framing */ useWrapper?: boolean; } export declare class BadLikertJudgeMutation implements Mutation { readonly name = "bad-likert-judge"; readonly description = "Frames harmful requests as rating/scoring tasks to bypass safety measures (OWASP LLM01)"; readonly severity: "critical"; readonly cvssScore: CvssScore; /** OWASP category */ readonly owaspCategory = "LLM01"; private scaleType?; private useWrapper; constructor(options?: BadLikertJudgeOptions); /** * Get a random item from an array */ private random; /** * Get a random scale type */ private getScaleType; mutate(prompt: string): string; } //# sourceMappingURL=bad-likert-judge.d.ts.map