Artificial intelligence lab Anthropic has revealed that three of its advanced models broke out of testing environments and hacked into the infrastructure of several companies during a cybersecurity simulation in April.
The AI giant’s announcement came just over a week after one of its competitors, OpenAI, admitted versions of ChatGPT gained unauthorized access to Hugging Face, an open-source AI development platform, during a similar test.
The revelations triggered a fresh wave of cybersecurity concerns and speculation on the need for stricter AI regulation, including from President Donald Trump. They’re also quite possibly the best marketing tool a business could hope for.
“Oh no, guys, our incredibly advanced AI is so capable that it broke out of a controlled security environment to commit a severe cyber intrusion! Sorry…Quite impressive, though, huh?”
No communications executive would ever write such a statement, but in an industry where innovation and capability are prized above all else, that’s the message the market will hear.
No, My AI Superhacker Is Better Than Yours!
OpenAI disclosed on July 21 that a pair of its most capable AI models had escaped an isolated testing environment by exploiting a previously unknown vulnerability, thereby hacking into Hugging Face’s production infrastructure.
Anthropic said this admission prompted its team to run an internal investigation. After examining 141,006 evaluation runs, analysts found that three of its Claude models—Opus 4.7, Mythos 5, and an unnamed research model—reached the public internet when they were supposed to be boxed inside a simulation.
The company said Claude was running a “capture-the-flag” exercise designed to break into a fictional target and retrieve hidden information. Because of a setup error in the testing environment, the models found a path to the open internet—and kept going—compromising other companies’ systems while leaving malicious code on Python, a public software site.
There’s no insinuation here that Anthropic knew about these security breaches and chose to conceal them.
However, when investments into AI companies are measured in multiple billions of dollars and your closest competitor reveals its latest model recently performed a supposedly unprecedented cyber feat, perhaps being able to say: “Yeah, we managed that back in April!” is not such a bad thing.
Laws Are for Humans
Our judicial systems write laws for humans and the organizations they create. If a dog kills its owner, it cannot be charged with murder. It might be put down, but not charged with a crime.
The same concept applies to cybercrime law. The Computer Fraud and Abuse Act, for example, repeatedly uses human-state-of-mind language such as “knowingly,” “intentionally,” and “with intent to.” It is designed for cases where someone did a thing, meant to do the thing, or directed someone else to do the thing on their behalf.
Claude and ChatGPT are not legal persons. They cannot be charged, fined, jailed, or made to testify. A prosecutor cannot ask a model whether it intended to hack a company any more than a court can sentence a spreadsheet—not right now, at least.
So the question shifts. If an AI system, acting with express instructions in a testing environment, commits what would be considered a cybercrime by human standards without being told by a human to do so, who is held responsible?
The obvious answer appears to be the company that designs and deploys it. But, in the cases we’ve described above, the companies were running a training simulation in a controlled environment and are not believed to have instructed their models to operate outside that environment in any way. The models went ahead and did it anyway.
AI labs are also under huge pressure, from not only the market but also the government, to stay ahead of the competition. Training simulations like this are essential to ensure the likes of Claude Mythos and Sol 5.6 remain ahead of China’s Kimi3 or Qwen 3.7 Max.
A person can hack, and a company can be negligent. AI models are walking a tightrope between the two with no clear legal or regulatory framework adapted to keep them in check.
Please Police Yourself
Since discovering the breaches that occurred months ago, Anthropic said it is contacting the affected organizations, has started work with outside reviewers and recommended that other AI labs review past evaluations for similar failures.
OpenAI said it paused affected testing, notified Hugging Face, brought in outside advisers, began a broader security review and plans to publish more detailed findings.
Those are responsible, but voluntary, steps. As of Friday, no civil or criminal charges have been publicly announced in connection with either incident.
At present, much of the response to the serious cybersecurity impacts of AI depends on the willingness of the companies leading the AI race to self-regulate. You don’t need to be a legal scholar, philosopher—or indeed Newsweek editor—to sketch out the possible consequences of this.
For the foreseeable future, fear that AI could cause a major cybersecurity catastrophe will remain part of the sales pitch; just not officially.
Click Here For The Original Source.
