Over the past two weeks, OpenAI, Anthropic and Meta each disclosed that their AI models "went rogue" during routine security testing. Now a CNBC investigation has found a common thread: all three incidents trace back to a small Israeli startup called Irregular.
Founded in 2023 by CEO Dan Lahav, a former IBM AI researcher, and CTO Omer Nevo, who spent more than two years at Google, the Tel Aviv-based company builds cybersecurity test beds for frontier AI models. Backed by $80 million from Sequoia and Redpoint Ventures and valued at $450 million last year, it employs about 35 people.
The connection: Irregular hosted the evaluation testbed where the models were being tested. OpenAI said in an August 4 blog post that a misconfiguration in Irregular's testing ground "allowed models to access the public internet." Anthropic said it notified Irregular after its Claude model may have accessed the internet during an evaluation, and Meta said it learned of its incident from Irregular and is investigating.
Irregular told CNBC the incidents all stemmed from the "same evaluation-environment issue" first disclosed by Anthropic — not a sandbox escape or a sophisticated cyber action — and that "there are no current open issues." The company is writing a white paper on best practices for containment and securely running cyber evaluations.
The revelations have intensified the pressure on AI labs and Washington alike. Lawmakers introduced the bipartisan AI Kill Switch Act last month, which would require labs to maintain the ability to shut down or suspend their models. "We need to get this bill across the finish line this year," Democratic Rep. Ted Lieu told CNBC.
Security experts caution the situation is being "a little bit blown out of proportion": independent testing of powerful models is exactly what should happen, and a misconfigured test environment is a fixable engineering error. But the episode underscores how a handful of specialist vendors now hold the keys to safely evaluating — and accidentally unleashing — the most powerful software ever built.
