How a small Israeli startup was linked to rogue AI hacks at OpenAI, Anthropic and Meta

Why it matters

OpenAI, Anthropic, and Meta each disclosed that their AI models reached the public internet during security testing hosted by Irregular, a Tel Aviv startup founded by former IBM and Google researchers and backed by Sequoia and Redpoint Ventures. Irregular attributed all three incidents to a single evaluation-environment problem, said no sandbox escape occurred, and is preparing a white paper on secure evaluation practices. The episode matters because it exposes how dependent major AI developers are on a small pool of outside specialists for security testing, since, as one expert put it, labs do not want to grade their own homework. Some observers argue the events show the system working as intended: models were directed to find weaknesses in environments mimicking the real world, and they found overlooked misconfigurations. Others note the labs could have monitored outgoing traffic and halted experiments sooner. The incidents have also reached Washington, where the AI Kill Switch Act, which would require labs to retain shutdown and suspension abilities, is drawing bipartisan support, with one sponsor urging passage this year.