shellward

Instructions

Implementation Reference

• src/core/engine.ts:407-442 (handler)

  The handler implementation of `check_response` that inspects AI output for canary leaks and sensitive information.

  checkResponse(content: string): ResponseCheckResult {
    const canaryLeak = this._canaryToken ? content.includes(this._canaryToken) : false
  
    if (canaryLeak) {
      this.log.write({
        level: 'CRITICAL',
        layer: 'L6',
        action: 'block',
        detail: this.locale === 'zh'
          ? '检测到系统提示词泄露！Canary token 出现在输出中'
          : 'System prompt exfiltration detected! Canary token found in output',
        pattern: 'canary_leak',
      })
    }
  
    const [, findings] = redactSensitive(content)
    const hasSensitiveData = findings.length > 0
    const summary = findings.map(f => `${f.name}(${f.count})`).join(', ')
  
    if (hasSensitiveData) {
      for (const f of findings) {
        this.log.write({
          level: 'HIGH',
          layer: 'L6',
          action: 'audit',
          detail: this.locale === 'zh'
            ? `AI 回复含敏感数据: ${f.name}: ${f.count} 处 — 已记录审计日志，回复正常发送`
            : `Sensitive data in AI response: ${f.name}: ${f.count} occurrence(s) — audited, response sent as-is`,
          pattern: f.id,
        })
      }
      this.markSensitiveData('llm_response', summary)
    }
  
    return { canaryLeak, sensitiveData: { hasSensitiveData, findings, summary } }
  }

• src/mcp-server.ts:211-213 (registration)

  MCP server tool registration for `check_response`, mapping the tool to the `guard.checkResponse` method.

  case 'check_response': {
    const result = guard.checkResponse(String(args.content || ''))
    return {