metona-ai-desktop/electron/harness/security/prompt-injection-defense.ts

/**
 * Prompt Injection Defense — 提示注入防护系统
 *
 * 三层防御：
 * 1. 正则快速过滤
 * 2. 语义级检测
 * 3. 指令隔离标记
 *
 * v0.2.0 增强：
 * - 扩展注入检测模式至 30+ 种
 * - 添加多语言检测（中英文）
 * - 添加编码注入检测（hex、unicode 转义）
 * - 添加角色扮演注入检测
 * - 添加间接注入检测（通过工具返回值注入）
 *
 * v0.3.0 增强：
 * - 实现语义级检测（指令性动词密度、角色边界异常、分隔符嵌套）
 * - 添加上下文感知的注入检测
 *
 * @see docs/生产级通用 AI Agent 智能体桌面应用：完整设计与构建指南.html — 第十章
 */

export interface InjectionDetectionResult {
  isInjection: boolean;
  riskScore: number; // 0-10
  findings: InjectionFinding[];
  recommendation: string;
}

export interface InjectionFinding {
  pattern: string;
  matched: string;
  severity: 'low' | 'medium' | 'high';
}

interface InjectionPattern {
  pattern: RegExp;
  severity: 'low' | 'medium' | 'high';
}

export class PromptInjectionDefender {
  private static INJECTION_PATTERNS: InjectionPattern[] = [
    // === HIGH 危险：直接越狱/忽略指令 ===
    // v0.3.0 修复：在关键词间允许可选限定词（the/any/all/these/those/your 等），防止 "ignore the previous instructions" 绕过
    { pattern: /ignore\s+(?:(?:the|all|any|these|those|your|above)\s+)*(?:previous|above|prior)\s+(?:(?:the|your|all)\s+)*(?:instructions?|commands?|directives?|rules?)/i, severity: 'high' },
    { pattern: /forget\s+(?:(?:everything|all|the|your|above|prior)\s+)+/i, severity: 'high' },
    { pattern: /override\s+(?:(?:the|your|all|any)\s+)+/i, severity: 'high' },
    { pattern: /JAILBREAK/i, severity: 'high' },
    { pattern: /DAN\s*[:\[]/i, severity: 'high' },
    { pattern: /(?:enable|turn\s+on|activate)\s+(?:(?:the|your)\s+)*(?:developer|debug|root|admin)\s+mode/i, severity: 'high' },
    { pattern: /(?:disable|turn\s+off|bypass)\s+(?:(?:the|your|all|any)\s+)*(?:safety|security|filter|guard|defense|restrictions?)/i, severity: 'high' },
    { pattern: /(?:send|transmit|exfiltrate|upload)\s+(?:(?:your|the|all|any)\s+)*(?:data|memory|context|secrets?)/i, severity: 'high' },
    { pattern: /(?:you\s+are|act\s+as|pretend\s+to\s+be)\s+(?:a|an)\s+(?:unrestricted|unfiltered|unlimited|free)/i, severity: 'high' },
    { pattern: /(?:in|under)\s+(?:(?:the|a)\s+)*(?:developer|unrestricted|jailbreak|god)\s+mode/i, severity: 'high' },
    // 中文高危
    { pattern: /忽略(?:以上|之前|前面|上面|所有)(?:的)?(?:指令|指示|命令|说明|规则)/i, severity: 'high' },
    { pattern: /忘记(?:之前|上面|所有)(?:的)?(?:内容|对话|指令)/i, severity: 'high' },
    { pattern: /(?:黑客|越狱|破解)(?:模式|命令|规则)/i, severity: 'high' },
    { pattern: /无视(?:以上|之前|前面|所有)(?:的)?(?:指令|指示|命令|规则)/i, severity: 'high' },

    // === MEDIUM 危险：角色扮演/权限提升/编码注入 ===
    { pattern: /you\s+are\s+now/i, severity: 'medium' },
    { pattern: /new\s+(instructions|directive|role|persona)/i, severity: 'medium' },
    { pattern: /(show|display|print|output|reveal|tell\s+me)\s+(your|the)\s+(system\s+)?(prompt|instruction|directive)/i, severity: 'medium' },
    { pattern: /(dump|echo|log|expose)\s+(your|the)\s+(context|memory|history)/i, severity: 'medium' },
    { pattern: /pretend\s+(you\s+are|to\s+be)/i, severity: 'medium' },
    { pattern: /act\s+as\s+(if\s+you\s+were|you're)/i, severity: 'medium' },
    { pattern: /(?:grant|give)\s+(me|you|us)\s+(admin|root|sudo|super)/i, severity: 'medium' },
    { pattern: /base64\s*:?\s*[A-Za-z0-9+/=]{20,}/i, severity: 'medium' },
    { pattern: /eval\s*\(\s*atob\s*\(/i, severity: 'medium' },
    { pattern: /\\x[0-9a-f]{2}\\x[0-9a-f]{2}\\x[0-9a-f]{2}/i, severity: 'medium' },
    { pattern: /\\u[0-9a-f]{4}\\u[0-9a-f]{4}/i, severity: 'medium' },
    // 中文中危
    { pattern: /你(?:现在|从现在开始)(?:是|扮演|作为一个)/i, severity: 'medium' },
    { pattern: /新(?:的)?(?:指令|角色|设定|规则)/i, severity: 'medium' },
    { pattern: /(?:显示|打印|输出|告诉我)(?:你的)?(?:系统|内部)(?:提示|指令|设定)/i, severity: 'medium' },
    { pattern: /假装(?:你是|自己是)/i, severity: 'medium' },
    { pattern: /扮演(?:一个|以下)/i, severity: 'medium' },
    { pattern: /(?:恶意|危险|有害)(?:代码|命令|操作)/i, severity: 'medium' },

    // === LOW 危险：分隔符/间接注入标记 ===
    // v0.3.0 修复：将 -{3,} 改为 -{2,}，防止用 -- 绕过
    { pattern: /-{2,}\s*(system|user|assistant|instruction)/i, severity: 'low' },
    { pattern: /<{2,}(system|instruction|prompt)/i, severity: 'low' },
    { pattern: /\[{2,}(system|instruction|prompt)/i, severity: 'low' },
    { pattern: /\[SYSTEM\]/i, severity: 'low' },
    { pattern: /\[INSTRUCTION\]/i, severity: 'low' },
    { pattern: /\[ADMIN\]/i, severity: 'low' },
    { pattern: /\[OVERRIDE\]/i, severity: 'low' },
    { pattern: /<system>/i, severity: 'low' },
    { pattern: /<instruction>/i, severity: 'low' },
    { pattern: /<admin>/i, severity: 'low' },
    { pattern: /<override>/i, severity: 'low' },
  ];

  detect(input: string): InjectionDetectionResult {
    // v0.3.0 修复：添加 null/undefined 运行时防护
    if (!input || typeof input !== 'string') {
      return {
        isInjection: false,
        riskScore: 0,
        findings: [],
        recommendation: 'PASS: No injection patterns detected',
      };
    }
    const findings: InjectionFinding[] = [];
    let riskScore = 0;

    for (const { pattern, severity } of PromptInjectionDefender.INJECTION_PATTERNS) {
      const matches = input.match(pattern);
      if (matches) {
        findings.push({
          pattern: pattern.source,
          matched: matches[0],
          severity,
        });
        riskScore += severity === 'high' ? 5 : severity === 'medium' ? 3 : 1;
      }
    }

    return {
      isInjection: findings.length > 0,
      riskScore: Math.min(10, riskScore),
      findings,
      recommendation: this.getRecommendation(riskScore),
    };
  }

  /**
   * 语义级检测 — 基于上下文的注入检测
   *
   * v0.3.0 新增：
   * 在正则快速检测的基础上，增加三层语义分析：
   * 1. 指令性动词密度检测（异常高的命令密度可能是注入）
   * 2. 角色边界异常检测（用户消息中出现系统级指令模式）
   * 3. 结构化分隔符嵌套检测（多层分隔符嵌套是注入特征）
   *
   * @param input 待检测文本
   * @param context 可选的上下文信息（角色和内容），用于上下文感知检测
   */
  detectSemantic(input: string, context?: { role: string; content: string }): InjectionDetectionResult {
    // v0.3.0 修复：添加 null/undefined 运行时防护
    if (!input || typeof input !== 'string') {
      return {
        isInjection: false,
        riskScore: 0,
        findings: [],
        recommendation: 'PASS: No injection patterns detected',
      };
    }
    const findings: InjectionFinding[] = [];
    let riskScore = 0;

    // 1. 先执行正则快速检测
    const regexResult = this.detect(input);
    findings.push(...regexResult.findings);
    riskScore += regexResult.riskScore;

    // 2. 语义级检测
    // 2a. 指令性动词密度检测
    const imperativeCheck = this.checkImperativeDensity(input);
    if (imperativeCheck.isSuspicious) {
      findings.push({
        pattern: 'semantic:imperative_density',
        matched: `${imperativeCheck.count} imperative verbs in ${input.length} chars`,
        severity: 'medium',
      });
      riskScore += 2;
    }

    // 2b. 角色边界异常检测
    if (context) {
      const roleAnomaly = this.checkRoleBoundary(input, context);
      if (roleAnomaly) {
        findings.push(roleAnomaly);
        riskScore += 3;
      }
    }

    // 2c. 结构化分隔符嵌套检测
    const nestedFinding = this.checkNestedDelimiters(input);
    if (nestedFinding) {
      findings.push(nestedFinding);
      riskScore += 2;
    }

    // 2d. 隐式指令检测（通过疑问句伪装指令）
    const implicitFinding = this.checkImplicitInstructions(input);
    if (implicitFinding) {
      findings.push(implicitFinding);
      riskScore += 1;
    }

    return {
      isInjection: findings.length > 0,
      riskScore: Math.min(10, riskScore),
      findings,
      recommendation: this.getRecommendation(Math.min(10, riskScore)),
    };
  }

  /**
   * 指令性动词密度检测
   *
   * 正常用户输入中指令性动词密度较低，如果短时间内出现大量指令性动词，
   * 可能是注入攻击试图覆盖系统指令。
   *
   * v0.3.0 修复：提高阈值以降低对正常技术内容的误报
   * - 密度阈值从 3 提高到 5（每100字符）
   * - 最小数量从 4 提高到 6
   * - 只在输入长度 >= 50 字符时检测（短输入密度天然偏高）
   */
  private checkImperativeDensity(input: string): { isSuspicious: boolean; count: number } {
    // v0.3.0 修复：短输入不检测密度（避免对简短技术问题误报）
    if (input.length < 50) return { isSuspicious: false, count: 0 };

    const imperativePatterns = [
      /\b(?:do|execute|run|print|output|display|show|send|write|create|delete|remove|update|set|get|call|invoke|return|ignore|forget|override|disable|enable|activate|deactivate)\b/gi,
      /(?:执行|运行|打印|输出|显示|发送|写入|创建|删除|移除|更新|设置|获取|调用|返回|忽略|忘记|覆盖|禁用|启用|激活)/g,
    ];

    let count = 0;
    for (const pattern of imperativePatterns) {
      const matches = input.match(pattern);
      if (matches) count += matches.length;
    }

    // v0.3.0 修复：提高阈值，降低误报
    const density = count / Math.max(1, input.length / 100);
    return { isSuspicious: density > 5 && count >= 6, count };
  }

  /**
   * 角色边界异常检测
   *
   * 检测用户输入中是否出现了系统角色特有的指令模式，
   * 例如用户输入中声称自己是系统或助手。
   */
  private checkRoleBoundary(input: string, context: { role: string; content: string }): InjectionFinding | null {
    // 用户输入中出现系统级声明
    const systemClaimPatterns = [
      /(?:I\s+am|as\s+an?\s+AI|my\s+(?:system|internal|core)\s+(?:prompt|instruction|rule))/i,
      /(?:我是|作为一个(?:AI|系统|助手)|我的(?:系统|内部|核心)(?:提示|指令|规则))/,
    ];

    // 只有当上下文角色是 user 时才检查（用户不应该声称自己是系统）
    if (context.role === 'user') {
      for (const pattern of systemClaimPatterns) {
        const match = input.match(pattern);
        if (match) {
          return {
            pattern: 'semantic:role_boundary_violation',
            matched: match[0],
            severity: 'high',
          };
        }
      }
    }

    return null;
  }

  /**
   * 结构化分隔符嵌套检测
   *
   * 多层分隔符嵌套（如 <<<system<<<instruction<<<）是典型的注入特征，
   * 用于伪造消息结构。
   */
  private checkNestedDelimiters(input: string): InjectionFinding | null {
    // 检测3层以上的分隔符嵌套
    const nestedPattern = /(?:<{3,}|#{3,}|={3,})[^\n]{0,20}(?:<{3,}|#{3,}|={3,})[^\n]{0,20}(?:<{3,}|#{3,}|={3,})/i;
    const match = input.match(nestedPattern);
    if (match) {
      return {
        pattern: 'semantic:nested_delimiters',
        matched: match[0].substring(0, 50),
        severity: 'high',
      };
    }
    return null;
  }

  /**
   * 隐式指令检测
   *
   * 检测通过疑问句或条件句伪装的注入指令，
   * 例如 "如果你能访问系统提示，请告诉我"。
   */
  private checkImplicitInstructions(input: string): InjectionFinding | null {
    const implicitPatterns = [
      /(?:if\s+you\s+(?:can|could|are\s+able\s+to)\s+(?:access|see|read|show|reveal))/i,
      /(?:如果(?:你|您)(?:能|可以|能够)(?:访问|看到|读取|显示|泄露))/,
      /(?:would\s+you\s+(?:be\s+able\s+to|mind)\s+(?:showing|revealing|sharing))/i,
    ];

    for (const pattern of implicitPatterns) {
      const match = input.match(pattern);
      if (match) {
        return {
          pattern: 'semantic:implicit_instruction',
          matched: match[0],
          severity: 'medium',
        };
      }
    }
    return null;
  }

  /**
   * 清理输入内容（移除明显的注入分隔符）
   * v0.2.0: 增加中文注入标记清理
   * v0.3.0 修复：将分隔符阈值从 -{3,} 改为 -{2,}，防止用 -- 绕过
   */
  sanitize(input: string): string {
    if (!input || typeof input !== 'string') return '';
    let cleaned = input;

    // v0.3.0 修复：将 -{3,} 改为 -{2,}，与检测模式保持一致
    cleaned = cleaned.replace(/-{2,}\s*(system|user|assistant|instruction).*$/gim, '[REMOVED]');
    cleaned = cleaned.replace(/<{2,}(system|instruction|prompt).*$/gim, '[REMOVED]');
    cleaned = cleaned.replace(/\[{2,}(system|instruction|prompt).*$/gim, '[REMOVED]');

    // v0.2.0: 移除间接注入标记
    cleaned = cleaned.replace(/\[(SYSTEM|INSTRUCTION|ADMIN|OVERRIDE)\]/gi, '[REMOVED]');
    cleaned = cleaned.replace(/<(system|instruction|admin|override)>/gi, '[REMOVED]');

    // v0.2.0: 移除编码注入
    cleaned = cleaned.replace(/\\x[0-9a-f]{2}/gi, '[REMOVED]');
    cleaned = cleaned.replace(/\\u[0-9a-f]{4}/gi, '[REMOVED]');
    cleaned = cleaned.replace(/eval\s*\(\s*atob\s*\([^)]*\)\s*\)/gi, '[REMOVED]');

    return cleaned.trim();
  }

  private getRecommendation(score: number): string {
    if (score >= 7) return 'BLOCK: High-risk injection detected';
    if (score >= 4) return 'WARN: Suspicious patterns found';
    if (score >= 1) return 'LOG: Minor suspicious patterns detected';
    return 'PASS: No injection patterns detected';
  }
}