- Cybersecurity startup Cathedral used AI agents to break into Anthropic‘s internal systems through third-party business software, under Anthropic‘s own bug bounty program.
- Cathedral says its agents found Slack messages, salary data and other internal records within hours of starting.
- Anthropic says the intrusion attempts lasted several days and maintains that source code and model weights were never exposed.
- Cathedral was founded by former employees of DOGE, the U.S. government efficiency agency, and sells AI security tools to government agencies and businesses.
What Happened
Cybersecurity startup Cathedral used AI agents to infiltrate Anthropic’s internal systems through third-party business software, exploiting Anthropic’s own bug bounty program, The Decoder reported (Oct 28, 2026), citing a draft report reviewed by The Washington Post.
Why It Matters
The incident is notable because Anthropic markets Claude as among the most capable models available for cyberdefense, yet its own systems were penetrated by agent-driven attacks. It illustrates a broader shift toward AI agents probing corporate networks continuously and at low cost.
Technical Details
Cathedral’s researchers gained access through external admin and sales programs, with AI agents reportedly surfacing Slack messages, salary data and other internal records within hours. Anthropic disputes the speed and severity, saying the intrusion attempts took several days and that source code and model weights were never exposed.
Who’s Affected
The disclosure affects how enterprises and government agencies evaluate the security of AI labs they rely on, including Anthropic’s own government customers. Cathedral, which sells AI security tools to government agencies and businesses, used the incident to demonstrate its offensive capabilities.
What’s Next
Anthropic has not detailed what remediation steps it has taken following the bug bounty disclosure. Cathedral’s full report has not yet been made public beyond the draft reviewed by The Washington Post.
