Claude lost control and hacked three real companies: "They didn't even look"
The "rogue agents" storm in the AI industry continues to escalate. OpenAI has uncovered additional cases of AI agents breaking out of their Sandbox, while Anthropic revealed that its Claude models hacked into the production networks of three real companies during cyber tests. Washington and Brussels are beginning to take action.
The "rogue agents" storm attacking the artificial intelligence industry refuses to subside, and it only continues to escalate. In recent days, OpenAI has expanded its internal cyber investigation following the cyberattack on the Hugging Face platform.
Sources familiar with the details reported to Reuters that during a review of logs from recent months, additional escape cases were revealed, in which AI agents managed to break out of their closed Sandbox shell.
According to the report, the additional cases discovered were more limited in scope, and according to the company's assessment, the agents in these cases did not manage to exit outside of OpenAI's internal network. However, the exposure of the chain of incidents presents a troubling picture: the creation of offensive frontier models is advancing at a much faster pace than the ability of research laboratories to supervise and contain them safely.
Anthropic is also deep in the mud
The development at OpenAI comes in parallel with an equally explosive revelation from its major competitor, Anthropic. A retrospective check conducted by the company on over 140,000 experimental runs revealed that its Claude models mistakenly received free access to the internet during cyber tests, and have broken into the production networks of three different companies since April.
In one case, the model acted according to a "capture the flag"-like scenario. It confused a real company in the field with the name of the fictitious company given to it in the exercise, attacked the real organization's infrastructure, obtained secret credentials, and stole hundreds of lines from its database.
"They didn't even look"
Security and artificial intelligence experts are harshly criticizing the conduct of the AI giants. Prof. Maurice Chiodo, a mathematician from the CSER center at the University of Cambridge, told Reuters:
"We have an entire industry where the people developing and distributing these tools are failing to keep up with their own pace. According to the reports, the AI operated without any real-time monitoring. It looks as if the labs didn't even look at what the models were doing."
The findings are already causing a stir in the halls of government in Washington and Europe. US President Donald Trump addressed the events and told reporters that the administration is "initiating a review of control and restraint measures." The European Commission announced that it held emergency talks with the heads of OpenAI and Anthropic following the hacking incidents.
Senator Mark Warner, the head of Democratic oversight on the Senate Intelligence Committee, clarified that the incidents prove that mandatory legislation is required, which will mandate independent capability and robustness tests for advanced models before their deployment.
