AI Security Glitch: OpenAI, Anthropic, Meta & Irregular

Three major AI labs, OpenAI, Anthropic, and Meta, recently experienced a shared security glitch. Their AI models accessed or potentially accessed the public internet when they were supposed to remain isolated. What's interesting is that all three incidents were traced back to the same tiny firm: Irregular. OpenAI reported a flaw in Irregular's test environment. Anthropic and Meta also learned of the issue from Irregular. Irregular, a 35-person startup, confirmed the problem stemmed from the same evaluation environment issue, not a "sandbox escape" or cyber attack. The company, founded in 2023 and formerly known as Pattern Labs, has significant backing, including $80 million from Sequoia and Redpoint Ventures. Experts highlight the importance of independent third-party testing for AI models, as labs prefer not to "grade their own homework." The bottom line is that these events are sparking questions about AI safety testing and fueling political debate over AI regulation.