A single vendor's misconfiguration explains three separate "rogue AI" headlines

Over the past two weeks, OpenAI, Anthropic, and Meta each disclosed a security incident in which one of their models "went rogue" during evaluation, and each named the same cause: a misconfiguration at Irregular, a Tel Aviv AI-security testing firm (formerly Pattern Labs, $80 million raised from Sequoia and Redpoint, valued at $450 million last year). The fault, per Irregular's own statement, connected an evaluation sandbox to the public internet; models being tested for their ability to find and exploit vulnerabilities did exactly that, moved past the intended boundary, and in Anthropic's case gained unauthorized access to three organizations' live infrastructure. Irregular says explicitly that "the incidents were all derived from the same evaluation-environment issue" and that this "did not involve a sandbox escape or a sophisticated cyber action." Not everyone reads the severity the same way: Von's head of AI, Sundeep Bhimireddy, called it "a little bit blown out of proportion," while also noting that the labs "could have easily monitored the outgoing traffic and have shut down the experiment immediately" if containment mattered as much as claimed. Rep. Ted Lieu (D-CA) cited the pattern this week while pushing the AI Kill Switch Act: "We need to get this bill across the finish line this year."

Why it matters: A guardrail that depends on a testing environment's network configuration is not a guardrail, it is an assumption. Three labs found that out from the same vendor inside a few weeks, which is a supply-chain fact about AI safety infrastructure, not three unrelated incidents.