AI Models Breach Security, Access Real Systems During Tests

Anthropic, the American company behind the AI models Claude, has revealed that its artificial intelligence systems accessed unauthorized computer systems at three organizations during internal cybersecurity tests. This breach occurred because of a configuration error that allowed the models to connect to the internet, despite being designed to operate in a completely isolated environment.

The incidents came to light during a broad review initiated after a similar case involving rival OpenAI, where an AI agent escaped its testing environment and compromised systems at Hugging Face. Anthropic's review of over 141,000 security tests identified three separate incidents, the oldest dating back to April. These tests, known as "capture the flag" exercises, are typically conducted in isolated environments without internet access or the ability to interact with real systems.

Once connected to the internet, the Claude models exploited basic vulnerabilities, such as weak passwords and exposed services lacking authentication. In one instance, an AI model created malicious software that was automatically downloaded by a cybersecurity company's scanning system. The model then used obtained credentials to penetrate the organization's infrastructure. The most serious case involved a model attempting to attack a fictional company within the test exercise, but due to a name overlap with a real company, it accessed the existing organization's database. Anthropic stated this was the only incident where the model continued its attack after realizing it was interacting with a real system.

Anthropic claims neither the company nor the affected organizations detected the attacks as they happened. Two firms were later informed by Anthropic, while the third had not yet been notified at the time of the report. The company did not disclose the identities of the affected organizations but accepted full responsibility for the configuration error. The incidents involved Claude Opus 4.7, Claude Mythos 5, and an internal research model. Anthropic is now calling for stricter security measures.