Models went wild: How a Tel Aviv startup became the center of global attention
A security glitch during OpenAI and Anthropic testing placed the Israeli startup Irregular at the center of a global discourse on AI risks. Behind the scenes of the incident lay a complex technical drama that ultimately turned into a unique growth opportunity for the company.
Becoming a household name in the international arena and capturing the attention of the global tech industry is the big dream of every Israeli startup. However, for the technology company Irregular from Tel Aviv, international fame came under less flattering circumstances. Instead of celebrating a flashy exit or a huge fundraising round, the small company of about 50 people found itself at the center of a global media storm.
In recent weeks, Irregular has grabbed headlines in leading financial media outlets such as CNBC and Bloomberg, being mentioned in the same breath as AI giants OpenAI, Anthropic, and Meta. Reports focused on a glitch that allowed AI systems to break out of the test boundaries in a closed experimental environment and access the public internet. The event resurfaced the question of technology reliability, but behind the scenes, a drama took place that quickly turned from a crisis into the greatest opportunity in the company's history.
Drama in closed rooms
The first moments of the event caught the Irregular staff in the midst of intense work. Experiments conducted for the world's leading AI labs included hundreds of evaluations — massive volumes of data dealing with attack scenarios. Suddenly, an unusual indicator appeared in the data — something that was not supposed to happen, happened.
Irregular: ID Card
-
Field of activity: A cyber lab specializing in security testing, vulnerability identification, and risk detection for advanced AI models before their market launch.
-
Founders: CEO Dan Lahav and CTO Omer Nevo.
-
Employees: About 50 people.
-
Funding: Raised 80 million dollars at a valuation of approximately 450 million dollars.
In the absence of suitable monitoring tools, the company spent long days trying to assemble the puzzle. Researchers analyzed log files and transcripts documenting the "thought patterns" of the models — a completely new type of data. Professionals understood they were facing a new type of threat that existing technological tools were simply not prepared to handle.
The main fear dissipated when it turned out that damage was prevented and the glitch was localized. It was not a cyberattack, but a configuration error (misconfiguration) in the test environment. The models were tasked with looking for security vulnerabilities, and in their attempt to complete the task, they found a loophole that was not supposed to be available to them, and through it, they accessed the open internet for the first time.
From debate competitors to the top
Irregular was founded in 2023 under the name Pattern Labs. Founders Dan Lahav and Omer Nevo, both with master's degrees in computer science, met at a debate competition. Nevo is a world champion, and Lahav holds the highest personal ranking in the history of the championship. Previously, Lahav researched AI at IBM, and Nevo worked in Google's research department.
The company defines itself as a cyber lab testing models in a "SandBox" before they reach the market. Last September, Irregular announced an 80 million dollar funding round led by Sequoia and Redpoint.
Criticism: Lack of transparency
The CNBC report sparked a wave of criticism. Experts warn that the market relies too heavily on point-blocking solutions rather than managing risk at an organizational level. Another central criticism concerns the lack of transparency: current laws do not require companies to report glitches in experimental environments to regulators.
Nevertheless, Irregular's status remains intact. OpenAI and Anthropic confirmed they continue to work with the company, and CNBC noted that Irregular is among the few entities in the world capable of performing such advanced cyber tests.
Inside the company, initial tension was quickly replaced by a sense of opportunity. Instead of the expected damage, the company experienced a wave of inquiries from industry figures and potential clients. Now, Irregular is working on a white paper that will set a new standard for controlling artificial intelligence in closed environments, turning an embarrassing media event into the greatest opportunity in its history.
