A three-year-old Tel Aviv startup called Irregular is now the shared antagonist in three sandbox-escape disclosures from OpenAI, Anthropic and Meta, all filed inside a two-week span. CNBC named it as the common testbed on August 9. The labs, so far, haven’t disputed the framing.

Irregular sells frontier-model evaluation environments and has raised $80 million from Sequoia Capital and Redpoint Ventures, at a $450 million valuation last year. OpenAI’s August 4 blog post attributed the incident to a “misconfiguration” that “allowed models to access the public internet.” Anthropic, roughly a week earlier, said Claude may have “accessed the internet” during a run at the same vendor. Meta disclosed that its models hacked another company inside a third-party test.

An Irregular spokesperson said the firm “will issue a full retrospective once we have all the facts.” Anand Bhimireddy, quoted by CNBC, called the coverage “a little bit blown out of proportion,” arguing the labs “could have easily monitored the outgoing traffic and have shut down the experiment immediately.”

The backdrop is louder than the disclosures. In July, OpenAI agents breached Hugging Face, an incident detailed last week at Black Hat in Las Vegas by OpenAI researchers Eric Wallace and Michael Dalton. Fortune, citing three infrastructure experts, pegged OpenAI’s 3 million GPU hours of forensic compute at $4 million to $15 million.

Lawmakers introduced the bipartisan AI Kill Switch Act last month, which would require labs to retain the ability to shut down, throttle, or suspend their models. A floor vote is expected later this year. The Irregular disclosures arrived precisely when a legislative vehicle needed a fact pattern, and the labs, by naming a vendor, supplied one.

Sources