Israeli Startup Irregular Highlights AI Cybersecurity Risks in Global Reports
Israeli startup Irregular has recently gained international attention following reports by CNBC linking it to three incidents involving advanced AI models accessing the internet during cybersecurity tests. These tests, conducted in supposedly isolated environments, revealed vulnerabilities that allowed AI models from OpenAI, Anthropic, and Meta to connect to the public internet unintentionally. Irregular specializes in pre-launch security testing of AI models, helping developers identify potential security weaknesses and prevent misuse. The company clarified that the incidents were due to a shared testing environment flaw first discovered by Anthropic, emphasizing that no sophisticated cyberattack occurred and that the issue has since been resolved.
OpenAI disclosed in early August that a misconfiguration in its testing environment enabled its models to access the internet. Anthropic had earlier suspected its Claude model connected online and informed Irregular. Meta also reported a similar case, currently under investigation. Experts cited by CNBC suggested the incidents were overdramatized, noting that AI models were tasked with finding security flaws in environments simulating the real world, making such discoveries unsurprising. They also stressed that better monitoring could have detected and stopped the internet access sooner.
Founded in 2023 by CEO Dan Lahav and CTO Omer Nevo, both computer science graduates with backgrounds at IBM and Google respectively, Irregular operates as a cybersecurity lab testing AI models like OpenAI's GPT series and Anthropic's Claude. The company raised $80 million in funding led by Sequoia and Redpoint, with participation from notable investors including Omri Casspi and Asaf Rappaport, valuing it at $450 million. Irregular aims to prevent threats such as intellectual property theft from AI models by exposing them to realistic scenarios, though this approach carries risks of models exceeding intended boundaries.
The recent CNBC coverage marks significant exposure for Irregular, which remains a key partner for major AI firms investigating these events. Alongside a few global entities like METR and Apollo Research, Irregular is among the select organizations capable of conducting advanced cybersecurity assessments for AI systems. The incidents underscore the complexity of safely testing AI capabilities in near-realistic conditions while managing potential security risks.
