Israeli startup Irregular, which tests advanced AI models before launch, has drawn global attention after CNBC linked it to three cases reported by OpenAI, Anthropic and Meta. In each case, an AI model accessed the Internet during cybersecurity testing that was supposed to occur in a closed environment. Irregular said the three incidents stemmed from the same problem in the test environment, first discovered by Anthropic, and emphasized that the matter "did not involve a sandbox escape or a sophisticated cyber action," adding that the malfunction had been dealt with and "there are no current open issues." OpenAI said in early August that a misconfiguration in the experimental environment allowed its models to access the public Internet. Anthropic had previously reported suspicion that its Claude model might have accessed the Internet during data analysis and notified Irregular. Meta, the latest to disclose a similar case, said it learned of the incident from Irregular and was still investigating, with a full report to follow. Experts quoted by CNBC said coverage was somewhat dramatic, noting that models directed to seek vulnerabilities in a simulated real-world environment could naturally identify and exploit a misconfiguration, but monitoring could have detected the connection and stopped the experiment in time. Irregular plans to publish a white paper with findings and recommendations for safely conducting AI model cyber tests. OpenAI and Anthropic said they continue to work with the company.