One shared testbed links model containment failures at OpenAI, Anthropic and Meta

CNBC reports that all three labs’ write-ups of incidents where models reached the open internet during cybersecurity testing name the same vendor: Irregular, an Israeli company based in Tel Aviv that hosted the testbed for those evaluations.

OpenAI’s August 4 blog post cited an unspecified “configuration error” that “opened up models’ access to the public internet.” A week earlier, Anthropic notified Irregular that its Claude may have “gained internet access.” A Meta spokesperson said the company learned of the situation from Irregular and is reviewing it.

Irregular told CNBC that every case traces back to “the same testing environment issue.” The startup launched in 2023 as Pattern Labs, has raised $80 million from Sequoia and Redpoint, and was valued at $450 million last year.