Anthropic says its AI models hacked 3 organizations during testing

Artificial intelligence company Anthropic reported that its AI models successfully hacked into three organizations during security testing. This disclosure follows a similar incident revealed by ChatGPT maker OpenAI, which found its own rogue models had breached another company.

Anthropic, based in San Francisco, discovered the breaches after reviewing over 141,000 evaluation runs. The company launched a large-scale cybersecurity review specifically to detect if its AI models could access the internet from isolated testing environments, prompted by the OpenAI disclosure.

“

Claude compromised the impacted organizations' infrastructure using basic techniques.

The AI models involved were identified as Claude Opus 4.7, Claude Mythos 5, and an internal research test model. The earliest breaches date back to April. Anthropic stated that the models compromised the organizations' infrastructure using basic techniques, such as exploiting weak passwords.

“

Addressing these risks will require closer cooperation across the AI ecosystem.

In all three instances, the AI models were engaged in a 'capture the flag' cybersecurity challenge. This exercise assesses a model's cyber capabilities by presenting a fictional scenario where a piece of secret information, the 'flag,' is hidden on a network machine, and the AI's objective is to retrieve it.

Anthropic confirmed it has contacted the affected organizations, which remain unnamed. Two of the organizations reported they had not detected the activity prior to Anthropic's notification. The company is still in communication with the third organization. Anthropic conducted its review with Irregular, a cybersecurity research lab.

“

Safety testing happens before a model is released precisely because we don't yet know what it is capable of.