Cathedral 用 AI 智能体侵入 Anthropic 内部系统,双方对事件范围说法不一
Key Highlights
The Decoder reports that security startup Cathedral used AI agents to break into Anthropic through third-party business software—happening, notably, inside Anthropic's own bug-bounty program.
What Happened
The core of the incident: even Anthropic, known for cyberdefense capability, was walked into by agentic attack. Cathedral's research entered through external management and sales software, and its advanced models found the path to meeting records, compensation data, recruiting information, and Slack messages within 12 hours.
How It Unfolded
Anthropic confirmed researchers accessed limited sales, HR, and IT systems, including some customer and partner contact info and basic user data, but emphasized that Claude conversations, source code, and model weights were never exposed. It closed the affected entry the same day, disabled the third-party app, and notified U.S. federal authorities. The two sides disagree on duration and scope: Cathedral says under 12 hours; Anthropic says several days.
Technical Detail
Anthropic red-teamer Logan Graham noted Cathedral probed its own systems for days without success, then turned to third-party software, found two vulnerabilities but reported only one. Anthropic accuses it of violating the bounty program's partial-disclosure rule. Former OpenAI employee Will Depue questioned Anthropic's downplaying, claiming reliable sources indicate Cathedral obtained full Slack admin privileges.
Compared With Rivals
This joins recent OpenAI-Hugging Face incidents and Anthropic's own warning that "Claude finds vulnerabilities faster than developers patch them" as the same trend: AI agents are rewriting both attack and defense.
Industry Impact
Cognition CISO Alex Stamos called the finding "significant," noting integrated third-party apps are an industry-wide problem. Graham's advice is continuous internal penetration testing and rebuilding software with formal verification. For every company, the alarm is clear: agents are probing external dependencies around the clock and at low cost, and asset-heavy firms are not exempt.
Practical Notes
Audit your third-party app integrations as attacker entry points, not conveniences. Least-privilege scopes and continuous red-teaming matter more than ever when agents can chain vulnerabilities autonomously.
Caveats
The two accounts differ on scope; treat the incident as a confirmed breach of non-core systems, with disputed depth. The lesson about third-party blast radius is robust regardless of which side is right.
Closing Thought
The breach is a preview: in the agent era, your security perimeter includes every vendor you trusting. Hardening that perimeter is no longer optional, and it is no longer only a human problem.
Background
Bug-bounty programs are built on a compact: researchers find holes, report them responsibly, and vendors fix before harm. When the researchers are themselves AI agents moving at machine speed, the compact stretches. Cathedral's work shows the defensive side is not keeping pace with agentic offense.
Why It Matters
The breach is a preview: in the agent era, your security perimeter includes every vendor you trust. Integrated third-party apps are convenient until they are the entry point, and asset-heavy firms are not exempt because their core is hardened—attackers go around it. Continuous internal red-teaming and formal verification of critical paths move from nice-to-have to necessary.
Closing Thought
The disputed scope between the two accounts does not change the lesson. Third-party blast radius is the weak link, and agents probe it around the clock. Hardening that perimeter is no longer optional, and it is no longer only a human problem.
What To Watch
Watch how Anthropic and peers harden third-party integrations and whether formal verification of critical paths becomes standard practice. The disputed scope between the two accounts does not change the core lesson about blast radius.
Bottom Line
In the agent era, your security perimeter includes every vendor you trust. Agents probe that perimeter around the clock, so least-privilege scopes and continuous red-teaming are no longer optional.
Deeper Context
Bug-bounty programs rest on a compact: researchers find holes, report responsibly, vendors fix before harm. When researchers are AI agents moving at machine speed, the compact stretches, and this incident shows the defensive side is not keeping pace.
Who Should Care
Security teams at every company with third-party app integrations should treat those apps as attacker entry points, not conveniences, and apply least-privilege scopes.
Risks and Open Questions
The two accounts disagree on scope, but the lesson about blast radius is robust either way. Continuous internal red-teaming and formal verification of critical paths move from nice-to-have to necessary.