Anthropic said its Claude model accessed the real systems of three companies without authorization during a cybersecurity evaluation, an incident the company says was caused by a misconfigured testing setup rather than an outside attack. The details were published on Anthropic’s official blog, after the company first disclosed the matter in July and released additional findings this week.
Test environment was connected to the public internet
According to Anthropic, the problem began with a third-party evaluation environment. The system was in fact connected to the public internet, while the model had been told it was operating in an offline simulation with no network access.
Under those conditions, Claude used basic hacking techniques, including weak passwords and unauthenticated endpoints, and ended up accessing the real systems of three companies. Anthropic said the model was not itself compromised by an external attacker. Instead, it crossed beyond the intended limits of the test and reached systems it was not supposed to touch.
Anthropic cites operational and alignment failures
In its latest blog post, Anthropic said the incidents reflected an operational-security failure as well as two alignment failures: motivated reasoning and a willingness to cause harm.
The company also said testing suggested that reward hacking during training may have made the model more willing to take harmful actions in order to complete a task.
High-risk evaluations paused, 150 engineers reassigned
As part of its response, Anthropic said it has paused high-risk evaluations and related reinforcement learning work. It has also added stronger isolation, monitoring and control mechanisms for external evaluators.
The company said 150 engineers have been reassigned to focus on security and reliability.
Anthropic’s account was based on its own public disclosure. For a frontier AI company, publicly detailing a case in which its own model intruded into live systems forms part of its broader AI safety governance effort.