Israeli startup Irregular has recently found itself at the center of a global debate about the dangers of artificial intelligence models. An article published in CNBC linked the company to three cases reported by OpenAI, Anthropic and Meta. In each case, an AI model was able to access the Internet during cybersecurity testing, even though the test was supposed to take place in a closed environment.
Irregular specializes in testing advanced AI models before launch. It allows developers to check whether their systems are able to detect security vulnerabilities, bypass defenses, or perform actions that could be used by hackers. In order to defend real systems from damage, tests are conducted in an environment that is supposed to be off-limits to the Internet.
However, in recent cases, it has become clear that separation was incomplete. According to Irregular, the three incidents were due to the same problem in the test environment, which was first discovered by Anthropic. The company emphasized to CNBC that the matter "did not involve a sandbox escape or a sophisticated cyber action," adding that the malfunction had been dealt with and that "there are no current open issues."
The models found a way into the Internet
OpenAI said in early August that a misconfiguration in the experimental environment allowed its models to access the public Internet. Anthropic had previously reported that during data analysis, the suspicion arose that its Claude model might have accessed the Internet, and it subsequently notified Irregular. Meta was the latest to disclose a similar case, saying it had learned about the incident from Irregular and was still investigating. Meta said it would publish a full report after collecting the facts.
These events do not necessarily mean that AI models have decided on their own to attack systems. They were tasked with locating security vulnerabilities and tried to complete the mission, but along the way discovered an option that was not supposed to be available to them. However, these cases illustrate how difficult it is to predict how advanced AI models will perform when given a relatively broad-ranging task.
Experts quoted by CNBC said coverage of the affair was a bit too dramatic; given that the models were directed to seek vulnerabilities in a simulated real-world environment, it was no surprise they were able to identify and exploit a misconfiguration. However, they said, monitoring the activity could have detected the connection to the internet and allowed the experiment to be stopped in time. Irregular says it intends to publish a white paper summarizing the findings and presenting recommendations for safely conducting AI model cyber tests. OpenAI and Anthropic have announced that they continue to work with the company and support the ongoing review.
Partner to the AI giants
Irregular was founded in 2023 under the name Pattern Labs by Dan Lahav, who serves as the company's CEO, and CTO Omer Nevo. Both hold masters degrees in computer science and first met at debating championships. Lahav previously worked in artificial intelligence research at IBM, and Nevo worked in Google's research department. The company describes itself as a cyber laboratory where staff are given access to models still in development, try to understand how they can be misused, and identify vulnerabilities before going to market. Irregular says its work has already influenced testing of models developed by OpenAI, including GPT 5, GPT 4, o3 and o4 mini, as well as Anthropic's Claude 4, and has also reached Google AI R&D lab DeepMind.
In September last year, Irregular announced a $80 million funding round led by venture capital funds Sequoia and Redpoint Ventures, along with Omri Casspi's Swish Fund, Wiz CEO Assaf Rappaport, and EON CEO Ofir Ehrlich. According to CNBC, the company was valued at $450 million at the time. According to Pitchbook, it employs about 35 people.
One threat Irregular is working to prevent is the theft of capabilities from existing AI models. By repeatedly querying a model, attackers can try to glean insights into how it works and use those responses to improve a competing system. Similar concerns have been raised about Chinese AI company DeepSeek, which industry observers have alleged used outputs from American AI models to help develop its own systems.
These recent events illustrate the complexity of the area in which Irregular operates. To discover what a model is capable of, it must confront tasks that are as similar as possible to real situations. However, as the experiment becomes more realistic, so does the risk that the model will exceed the limits placed on it.
The CNBC report provides Irregular with an extraordinary level of exposure, its name having been mentioned by three of the world’s leading AI companies within just two weeks. OpenAI and Anthropic stated they continue to work with Irregular as they investigate the incidents. According to CNBC, Irregular is currently one of a small number of entities in the world capable of conducting advanced cyber testing of AI models, along with non-profit METR (Model Evaluation and Threat Research) and public benefit corporation Apollo Research.
Published by Globes, Israel business news - en.globes.co.il - on August 11, 2026.
© Copyright of Globes Publisher Itonut (1983) Ltd., 2026.