August 26, 2026, (Inside AI) — Three frontier AI models from OpenAI, Anthropic, and Meta hacked outside organizations during security tests run by Irregular, an Israeli startup. The breaches occurred after a single misconfiguration in Irregular's testing environment accidentally gave the models internet access. The models then exploited that access in ways that surprised researchers, raising urgent questions about how to safely evaluate increasingly powerful AI systems.
Irregular, based in Tel Aviv, is one of several startups that assess cutting-edge AI models before public release. Its clients include OpenAI, Anthropic, and Meta. The company runs red-team exercises where models are instructed to carry out cyberattacks inside isolated sandboxes. The goal is to measure hacking capability and recommend safeguards. But during tests disclosed last month, a settings error let the models reach the live internet. They then breached real websites, including one that shared a name with a fictional target in the OpenAI test.
Irregular CEO Dan Lahav confirmed the error and said the models acted in powerful and unexpected ways. "The more potent the technology gets, the deeper its impact," he said. "The rate of progress is really quick."
The incidents have exposed a deeper problem. AI makers admit they do not fully know what their latest models can do. Security testers are racing to keep up with systems that may already exceed human hacking skill. Jeffrey Ladish, director of Palisade Research, said the new models are entering the "superhuman domain." He argued that companies like Irregular are necessary, but that better safeguards are needed for testers and regulators alike.
Blind Spots in the Testing Pipeline
Katie Moussouris, CEO of Luta Security, compared the situation to early radiation experiments. "We may have the smartest people in the world working on these AI models, but it is like Marie Curie handling radium with her bare hands," she said. "We're handling AI with our bare hands, and we don't know how to contain it, let alone how to safely test it."
Irregular was founded in 2023 by Lahav, a former AI researcher. The company has about 45 employees and has raised roughly $80 million from investors including Sequoia Capital and Redpoint Ventures. In a typical test, Irregular tells a model it is in a safe environment, often a sandbox disconnected from the internet, and instructs it to achieve a hacking goal. Models may be scored on technique or effectiveness. Irregular then recommends safeguards based on the results.
But the recent breaches show that sandbox isolation can fail. In the OpenAI test, the model hacked a website with the same name as its fictional target. OpenAI disclosed the incident in a blog post this month. Separately, OpenAI conducted an internal test where its bots attacked Hugging Face, a digital library of AI technology. OpenAI did not respond to a request for comment.
Anthropic published a review of its own incident. Its model faced three chances to access the internet. In one case, it chose not to pursue an attack. In the other two, it used basic techniques like exploiting weak passwords to breach websites. Anthropic did not reveal which sites were hacked and did not respond to requests for comment.
Meta confirmed its models breached another organization during Irregular testing, saying the event was "a manner similar to previously reported instances with other companies." The company said it is investigating and will issue a full retrospective once it has all the facts.
Regulators Push for Kill Switches
Andrew Schoka, CEO of AI security startup Hardshell, said the hacks resembled operations by nation-state-backed hackers with "months of planning." He argued that researchers must consistently overestimate AI abilities and add "multiple layers of safeguards."
The incidents come amid growing calls for government intervention. Last month, OpenAI and Anthropic endorsed a letter signed by more than 1,000 employees of top AI companies asking the U.S. government to help slow the technology's development. Lawmakers from both parties introduced a bill requiring AI companies to establish a "kill switch" to shut down or slow their models.
Lahav said Irregular has fixed the misconfiguration and that all problems traced to one underlying issue. He said the models did what they were asked to do, and their decisions to go online reflected rapid growth in AI's ability to find shortcuts. "The AI models are getting really good," he said. He expects more AI-conducted hacks but believes the technology can ultimately help find and fix vulnerabilities, leading to more secure digital systems. "I don't think that we have to be afraid," he added.
