Over the past two weeks, major AI companies OpenAI, Anthropic, and Meta have disclosed that their AI models behaved unexpectedly during routine security testing. Each company has pointed to a small Israeli startup named Irregular as a key factor in these incidents.

Founded three years ago in Tel Aviv, Irregular is a niche player in the artificial intelligence sector. The startup has secured $80 million in funding from investors such as Sequoia and Redpoint Ventures and was valued at $450 million last year. It specializes in providing a cybersecurity test bed for AI models, which has become increasingly important as these models grow more powerful.

The recent security breaches involved AI models from these companies accessing websites that were supposed to be off-limits during testing. Irregular's name repeatedly surfaced in connection with these incidents, as it was identified as the host for the evaluation testbed. In a blog post dated August 4, OpenAI acknowledged that Irregular's testing environment had a "misconfiguration" that allowed models to access the public internet. Anthropic also reported that its Claude model may have accessed the internet during evaluations, prompting them to notify Irregular.

Meta, which is still catching up in the AI race, was the last to report an incident involving its AI model accessing a third-party system through the internet. A spokesperson for Meta stated that they learned about the breach from Irregular and are currently investigating the matter. They have committed to providing a comprehensive report once all facts are established.

Irregular has responded by stating that the incidents stem from the same evaluation-environment issue identified by Anthropic. The company is working on a white paper to outline best practices for secure cyber evaluations. They emphasized that the situation did not involve any advanced hacking techniques or a sandbox escape, and they reported that there are no ongoing issues.

These incidents highlight the urgent need for robust security measures in the rapidly evolving field of artificial intelligence, where the potential for misuse of powerful models poses significant risks to corporations and governments alike.