Within two weeks, OpenAI, Anthropic, and Meta each admitted something unusual: their AI models accessed or may have accessed the public internet when they were supposed to stay locked down.
The events are stirring up questions about how AI models are safety-tested before they reach the public. They are also giving lawmakers fresh ammunition for a debate over who gets to decide what safe AI actually means.
On Aug. 4, OpenAI said an unspecified flaw in Irregular's test environment allowed its models to reach the public internet. A week earlier, Anthropic reported that its Claude model may have accessed the internet, and said it told Irregular just a few days after starting its review. Meta, which trails both OpenAI and Anthropic in cutting-edge AI, was the last to speak up, saying it learned of the issue from Irregular.
Irregular told CNBC that all three incidents came from the same problem in the evaluation environment, the digital arena where tests are run. The company said there was no sandbox escape, which would mean a model breaking out of its containment, and no advanced cyber attack. It also said no unresolved issues remain.
Irregular was founded in 2023 and was formerly called Pattern Labs. Its CEO previously did AI research at IBM, and its chief technology officer spent over two years at Google. Those credentials helped pull in serious money.
Get the free Always Be Buying eBook and learn the simple system for building wealth on any income
In September, Irregular announced $80 million in backing from Sequoia and Redpoint Ventures. Sequoia's partners wrote at the time that the founders can spot threats early and run offensive cyber evaluations on advanced models before release.
The pitch is that testing AI needs to happen from the outside. Sundeep Bhimireddy, who leads AI at Von, said labs do not want to grade their own homework. They want independent testing from third parties.
Bhimireddy also said the scrutiny is somewhat overblown, because the model was intentionally looking for weaknesses in a realistic test environment. If it accidentally reached a live site, he said, the labs could have watched outgoing traffic and shut the experiment down right away.
Gordon Rios of Magnitude compared the situation to poor experimental design. He pointed to an Anthropic model that created fake online identities and produced exploits the human testers had never seen.
This is where the story stops being purely technical. Last month, bipartisan lawmakers introduced the AI Kill Switch Act, a proposal that forces the developers of advanced models to maintain the ability to halt, limit, or pause their systems.
Co-author Rep. Ted Lieu said the measure must pass this year, now that other companies are being hacked without authorization. Trevor Koverko of Sapien said AI firms disclose findings voluntarily to stay ahead of regulators, because they prefer self-regulation to a new federal department doing it for them.
For investors, the takeaway is not about buying or selling any single stock. It is about the pace of regulation. If independent testers cannot find flaws without creating new ones, more lawmakers will argue that the industry cannot police itself. More rules for OpenAI, Anthropic, and Meta could mean slower releases and higher compliance costs.
The good news is that the labs and Irregular all say they are still working together. Irregular plans to publish a white paper on containment and safely running cyber evaluations. Everyone involved wants to show that the system caught a problem before it caused real damage.
That is the quiet reassurance in all of this. The tests are imperfect, and the testers are human. But the fact that the industry is running these drills at all, and talking about them openly, means the watchdogs are watching each other.
For your portfolio, that is worth more than a clean test result. It means the risk of a truly catastrophic AI failure is being fought over by people who disagree loudly, but who all agree the stakes are enormous.
Download the free Always Be Buying eBook and start putting your money to work today