Key points
- Cathedral accessed employee details and internal messages in less than 12 hours.
- Anthropic confirmed the breach but stated core model weights remained secure.
- The incident was discovered via an external researcher vulnerability programme.
Cathedral, a cybersecurity startup, accessed sensitive internal systems at Anthropic in less than twelve hours, exposing employee information and internal communications, according to a draft report reviewed by The Washington Post.reported
Evidence: reported
AI startup hacks Anthropic, gains access to employee details, internal messagesaifleet-hn-ai
The claim
The researchers used AI agents to identify weaknesses in business software that Anthropic employs for administrative and sales tasks. These agents ran for under twelve hours before gaining access to internal data, including meeting transcripts and Slack messages. The exposed information reportedly included salary details, job offers and internal discussions regarding Anthropic's advanced AI models.
The response
Anthropic confirmed that the researchers accessed a limited set of its sales, human resources and IT systems. The company stated that the exposed data included business contact details for some customers and partners, along with basic user account records. Anthropic emphasised that the researchers could not access conversations between customers and its Claude chatbot, nor the source code or model weights behind its AI systems.
The bottom line for teams
Anthropic stated that its monitoring flagged some of Cathedral’s activity before the researchers reported it.reported The company shut off their access on the same day. It disabled the affected business software and restored it only after vendors confirmed that the security flaws had been addressed. Anthropic briefed federal officials about the incident.
Evidence: reported
AI startup hacks Anthropic, gains access to employee details, internal messagesaifleet-hn-ai
What the paper shows
The draft report examined how AI agents interact with third-party business software. It demonstrated that automated tools can find vulnerabilities in administrative and sales platforms within a short timeframe. The results showed access to non-core internal data, such as HR records and Slack messages. The stated limitation is that the attack did not reach core AI model weights or customer chat logs.
Why it matters
- Model evaluators can see how AI agents exploit third-party software vulnerabilities rather than attacking core models directly.
- The incident highlights the risk of administrative and sales tools serving as entry points for sensitive data exposure.
- Monitoring systems detected the activity, suggesting that automated detection may catch some AI-driven intrusions before full compromise.
- Alex Stamos pointed to weaknesses in how companies manage third-party software.
Topics: Anthropic, AI agents, cybersecurity, third-party risk, Cathedral
Updated
Then / now / next
Anthropic relied on third-party business software for administrative and sales tasks without this specific vulnerability being known.
The company has disabled the affected software and restored it only after vendors confirmed fixes.
Federal officials were briefed on the incident, with national security concerns raised by a senior Trump administration official.
Story timeline
- Then Anthropic relied on third-party business software for administrative and sales tasks without this specific vulnerability being known.
- 31 Oct 2026 AI startup hacks Anthropic, gains access to employee details, internal messages
- Now The company has disabled the affected software and restored it only after vendors confirmed fixes.
- Next Federal officials were briefed on the incident, with national security concerns raised by a senior Trump administration official.
For readers
Researchers
- The report describes an incident where AI agents breached administrative software in under twelve hours.
- Core model weights remained secure during this specific test.
Policy
- Federal officials were briefed on the incident.
- National security concerns were raised by a senior administration official.
Business
- Third-party software vulnerabilities exposed employee and customer contact data.
- Vendors had to confirm fixes before systems were restored.
Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
Did the attackers access Anthropic's AI model weights?
No, Anthropic stated the researchers could not access the source code or model weights behind its AI systems.
How long did it take to breach the systems?
The AI agents accessed the internal systems in less than twelve hours.
What data was exposed?
Employee details, salary information, job offers, meeting transcripts and internal Slack messages were accessed.
Connected stories
Reading receipt
- Cathedral accessed employee details and internal messages in less than 12 hours.
- Anthropic confirmed the breach but stated core model weights remained secure.
- The incident was discovered via an external researcher vulnerability programme.