Anthropic systems breached in 12 hours by AI security agents

Anthropic confirmed unauthorized access to internal communication and operational tools on Wednesday after automated agents deployed by cybersecurity startup Cathedral breached systems in under 12 hours.

The intrusion, detailed in a draft assessment, exposed meeting transcripts, payroll records, and internal technical discussions regarding frontier artificial intelligence models, triggering immediate national security reviews across federal agencies.

Federal scrutiny widens over autonomous cyber penetration

Anthropic said in a statement that Cathedral gained limited access to sales, human resources, and information technology environments, adding that management notified federal authorities directly. The penetration, conducted by former Department of Government Efficiency technical personnel operating at Cathedral, demonstrated how automated software agents can locate lateral access points across commercial enterprise software without human direction.

US officials reviewing the findings said the speed of the automated operation raised significant concerns regarding foreign adversaries deploying identical agent-driven techniques against sensitive domestic infrastructure. The incident forced corporate security teams to re-evaluate defensive perimeters around frontier model training data and proprietary developer communications.

Federal cybersecurity regulators are scheduled to receive a completed incident assessment report from technical investigators by Nov. 6, 2026, to determine whether new disclosure rules are required for autonomous security testing.