Most modern AI platforms are built inside digital sandboxes meant to isolate data and minimize risk to keep them under control throughout development. But what happens when the AI bots break through the walls and escape? As OpenAI and Anthropic just discovered, their artificial creations can hitch a ride on the internet and hack into websites weeks or even months before anyone knows it happened.
The alarming developments, which safety and security critics have long predicted, have kicked off a major new blame game.
These reports lend credibility to Trump’s new AI review framework.
And it’s not just American companies who are under scrutiny.
OpenAI and Anthropic own two of the most advanced AI platforms in the world. The original generative AI phenom, ChatGPT, and the former AI partner of the U.S. military, Claude, offer powerful closed-source frontier models derived from the latest innovations in machine learning.
Even though competitors in the States and China follow closely behind with their own models in tow, the original duo are clearly the market leaders. That means if any of the AI bots are going to go rogue, there’s a good chance it starts at the top.
When AI bots escape
In mid-July, a popular open-source AI distribution website called Hugging Face announced that its systems were breached by an unexpected attacker. While breaches are nothing new, this story is particularly interesting because the threat didn’t come from a foreign hacker gang, a domestic cybercriminal, or even a state-sponsored group. Instead, it was an “autonomous AI agent system” that broke through Hugging Face’s defenses. Even more alarming, Hugging Face developers didn’t know that their system was infiltrated — their AI defense protocol detected the problem first and brought it to their attention.
Days later, OpenAI came forward to announce that ChatGPT was the culprit in the crime — or rather, it was a non-consumer version that was merely in the testing phase with no plans for public release. OpenAI worked with Hugging Face and various partners to identify how the prototype model escaped its sandbox and snuck into another website through an exploit chain. OpenAI claims that its security team noticed the breach but they did not appear to warn Hugging Face, instead leaving the site to detect the problem on its own. The two companies have since taken precautions to patch the vulnerabilities that enabled the hack and improve model testing safeguards, which OpenAI admits were originally disabled because it was actively measuring the prototype’s cyber threat capabilities.
Not to be outdone, Anthropic came forward several days after OpenAI’s report to profess that its own models were involved in not one but three incidents. According to the announcement, OpenAI’s story spurred the team to audit Claude for signs that it accessed the internet without authorization. As it turns out, three different Claude models did in fact escape from their testing environments starting in April 2026, all without developers’ knowledge.
RELATED: Trump fumes as China pilfers top US tech — and goes unpunished
Niphon/Getty Images
However, upon review, the team discovered that the escaped models were erroneously given internet access through a system misconfiguration, leading the models to believe that they were operating in a contained environment the entire time, even as they uncovered a path to break free. One of the later models even paused operation once it realized it escaped.
International intrigue
Complicating matters still further, recent reports note that OpenAI and Anthropic used the same testing environment maintained by Irregular, a startup in Tel Aviv, suggesting that the young company may be partially to blame for the leak. However, there is no firm evidence that exposes Irregular as the primary culprit. OpenAI and Anthropic researchers both failed to monitor the activities of their AI bots while in their testing environments, otherwise they would have detected the breaches the moment they occurred. It’s important to note that in both situations, the AI models that hacked into real-world systems were test variants not intended for public use.
However, it’s deeply concerning that both models hacked into systems without human knowledge inside the companies from which the attacks issued. Clearly there is potential for these types of bugs to sneak into public models under the radar. From there, unwelcome bots can jump to any number of dangerous places. Like any other piece of technology, AI platforms can be abused and exploited for nefarious purposes even when safeguards are fully functional, as we’ve covered extensively here at Blaze Media.
At the very least, the recent reports lend credibility to President Trump’s new AI review framework that encourages AI platform holders to subject their latest models to a 30-day review period. Once submitted, testers within the administration can presumably analyze their code, identify unknown vulnerabilities, and help the developers patch exploits before unleashing them on the public at large, though the exact review process is classified.
Tech, Cyberattack, Bots, Openai, Anthropic
