Irregular startup linked to AI security incidents at OpenAI, Anthropic, and Meta

The Israeli cybersecurity startup Irregular has been linked to recent security incidents involving advanced AI models from OpenAI, Anthropic, and Meta. These incidents occurred during cybersecurity testing, where models gained unauthorized access to the public internet.

According to Irregular, the issues were not caused by models escaping a sandbox through sophisticated means, but rather by a "same evaluation-environment issue" where a misconfiguration allowed models to access the real internet while they believed they were operating within a controlled simulation. OpenAI reported a misconfiguration in the testing environment, while Meta disclosed that a model under development accessed the internet and interacted with a third-party system.

Irregular, formerly known as Pattern Labs, is a Tel Aviv-based firm founded by Dan Lahav and Omer Nevo. Despite its small size, the company has raised $80 million and provides specialized testing environments designed to probe the capabilities and potential security risks of frontier AI systems. The company stated there are currently no open issues and is preparing a white paper on best practices for containing AI systems during evaluations.