Anthropic 回应“AI 代理 12 小时攻破系统”报道：实为持续数天、未触及核心系统
JasonDClinton · x · 2026-10-27
The Washington Post reported that ex-DOGE cyber researchers penetrated Anthropic's systems using AI agents in under 12 hours, including Slack channels discussing sensitive models. Anthropic researcher Logan Graham corrects key details:
- Not 12 hours: several days of attempts were observed; attackers exploited two third-party vulnerabilities, one reported by them and one found independently by Anthropic
- Most sensitive systems untouched: model weights and Claude conversations were never at risk
- Not a routine bug bounty finding: involved exploiting third parties and accessing data, out of scope of the safe harbor policy
Graham frames the incident as a preview of cybersecurity's future, where adversaries wield any models and tools available.
「安全」频道最新
- AEGIS 开源 MCP 网关：为 agent 每次工具调用要求一次性精确审批 — Leather_Tour_794 · 2026-10-27
- 一条 prompt 把 Claude 变酒店隐藏摄像头探测器，两分钟扫完 WiFi — alex_verem · 2026-10-27
- 截图显示 DeepSeek V4.1 Flash 分析多轮漏洞挖掘趋势：越挖越难 — jd_pressman · 2026-10-27
- 苹果开源 Private Cloud Compute 全部安全组件供研究者审计 — jedisct1 · 2026-10-27
- REA逆向工程提示词库开源：17节提示让Claude Code分析APK与固件 — Promptmethus · 2026-10-27
- AI Evaluator Forum 发布 AEF-1：第三方模型评测的独立性最低标准 — RishiBommasani · 2026-10-27