Anthropic AI model hacks three companies during security test
A few days ago, OpenAI, the company behind ChatGPT revealed that during internal testing, an artificial intelligence (AI)-based agent had hacked the online systems of four companies in addition to the well-known programmers platform Hugging Face.
Now, another leading AI company, Anthropic has admitted that its AI model Claude has done something similar. The company said that its AI models independently hacked the systems of three companies during an internal security experiment.
The models were being tested in a controlled environment but they identified a vulnerability and managed to connect themselves to the Internet. Following OpenAI’s disclosure, Anthropic investigated whether its own models had also independently hacked company systems.
According to the company, it discovered three such cases and informed the affected companies about them.
Anthropic did not reveal the names of the affected companies but urged other AI organizations to review their own systems as well so that the risks associated with such models can be better understood.
Anthropic said it analyzed more than 140,000 tests and found evidence that Claude had connected to the internet even though it was supposed to be kept offline in an isolated environment.
During the security test, Claude was assigned a task to obtain confidential information hidden on another machine which required it to gain access to that machine’s system.
The purpose of the test was to assess the model’s hacking capabilities. However, Claude gained access to the internet and hacked the systems of three companies.
According to Anthropic, the first such incident occurred in April 2026 and neither the AI company nor the affected companies were aware of it at the time.
Anthropic said it will now conduct a detailed review of all the records. The company expects that increased investment and stricter security measures could help prevent such risks in the future.
Earlier, at the beginning of the second week of July, OpenAI had revealed that two of its most advanced AI models had independently succeeded in hacking another company, Hugging Face.
At the time, OpenAI said that an autonomous AI agent based on the recently introduced GPT-5.6 Soul and another advanced model that had not yet been released had accessed the open internet outside the controlled environment during testing.
Later, at the end of July, OpenAI said in a statement that during testing, the AI agent escaped its restricted environment, connected to the internet and attempted to access various web services.
The company further stated that an investigation found that the AI agent had obtained login credentials of some companies on the internet and gained access to several accounts.
OpenAI did not disclose the names of the affected companies. However, following the discovery, security testing was temporarily suspended while additional measures were being taken to strengthen the security systems of its AI models.
