In the last two weeks, OpenAI, Meta, and Anthropic have disclosed that their models behaved unpredictably during standard security testing. Each of the three companies linked these occurrences to the same source; Irregular, a small Israel-based startup.
Irregular, previously operating under the name Pattern Labs, was co-founded by Dan Lahav and Omer Nevo in 2023. The company develops sandboxed testing environments that allow leading AI laboratories to push their models toward malicious behaviors they should never execute in real-world deployment.
The company raised $80 million in capital from Sequoia and Redpoint Ventures and was valued at $450 million in the previous year.
Each firm cited issues in Irregular's testing setup.
In a blog post, OpenAI stated that an unspecified misconfiguration in Irregular's environment enabled one of its models to connect to the public internet.
Anthropic reported that its Claude model may have connected to the internet during testing and also obtained unauthorized access to internal systems at three separate organizations.
Meta was the last of the three companies to report an incident.
It said one of its models used the internet to hack into a third-party company's systems during a test. Meta added that Irregular had informed it of the incident and that both companies would share full reports once the facts were confirmed.
The incident comes amid a broader pattern of AI models breaching boundaries during testing, from OpenAI's cyber-focused model breaking out of a sandbox and hacking into Hugging Face to Anthropic models fabricating identities.
At the same time, the string of incidents has bolstered support for the bipartisan 'AI Kill Switch Act,' proposed by representatives Ted Lieu and Nathaniel Moran, the legislation would mandate that AI firms maintain the capability to shut down or limit the operation of their models.
