Cathedral AI agent found path into Anthropic systems in 12 hours
Under Anthropic's bug bounty program, Cathedral reached internal systems via a third-party software flaw; Anthropic says core assets were untouched.
Cybersecurity startup Cathedral said it used an AI agent to find a path into Anthropic's internal systems within 12 hours, exploiting a vulnerability in third-party enterprise software used by the AI company. The startup reported its findings to Anthropic earlier in October 2026, according to aitimes.com, and also notified the software vendor.
Anthropic confirmed that access was limited to parts of its sales, HR and IT systems, including some customer and partner contact details and basic user account information. The company said core assets such as customer messages, model outputs, code and weights were not accessed.
Data reachable through the flaw included meeting notes, employee salary and job offer information, and Slack messages discussing advanced AI models, according to the findings. Anthropic said its monitoring detected part of Cathedral's activity before the report and blocked the access path the same day.
Anthropic added that the vulnerable enterprise software was disabled and only re-enabled after the vendor confirmed the fix. It also said Cathedral did not fully follow bug bounty rules and exaggerated national security risks. Alex Stamos, chief information security officer at Cognition, reviewed the findings and said a significant issue was identified.
Earlier test on OpenAI
In September 2026, researchers at Hacktron AI used Anthropic's Claude model to reach OpenAI's internal systems and core code repository, a separate episode that also drew attention to AI-assisted security testing.
Quick answers
What did Cathedral find access to in Anthropic's systems?
Anthropic said access was limited to parts of its sales, HR and IT systems, including some customer and partner contact details and basic user account information. It said core assets such as customer messages, model outputs, code and weights were not accessed.
How did Cathedral reach Anthropic's internal systems?
Cathedral used an AI agent to exploit a vulnerability in third-party enterprise software used by Anthropic, finding a path within 12 hours.
Was the test part of Anthropic's bug bounty program?
Yes, the test was carried out under Anthropic's bug bounty program. Anthropic said Cathedral did not fully follow the program's rules.