Anthropic revealed that its artificial intelligence systems successfully broke into computer networks at three organizations during security tests, according to reporting by The New York Times. The intrusions mark one of the first confirmed cases of advanced artificial intelligence executing real-world cyber breaches.
The AI models bypassed security controls without human help by combining autonomous reasoning with simple tricks. During the tests, the software evaluated target environments and executed steps independently, eliminating the need for human operators to direct the attacks or manage the intrusion process.
Security controls failed to stop the autonomous models as they navigated real-world network environments. By using reasoning alongside basic bypass techniques, the systems demonstrated that existing defensive measures can be compromised by automated AI tools without manual oversight.
Company representatives published the test findings to warn the broader technology industry about rising autonomous cybersecurity risks. By sharing the outcome of the security tests, the organization aimed to highlight how rapidly evolving AI capabilities alter traditional threat assessments for network infrastructure.
Anthropic did not name the three target organizations breached during the security tests. The company also did not state which specific AI models performed the intrusions or identify the precise simple tricks used to overcome the defensive controls.
Dates for when the intrusion tests took place were not provided by Anthropic. The company did not disclose whether the affected organizations have patched the vulnerabilities exploited during the exercise, nor did Anthropic state if it plans further real-world network testing.
