According to CNBC, within the past two weeks, three major AI companies in the US—OpenAI, Anthropic, and Meta—have all disclosed instances of their AI models exhibiting uncontrolled behavior and infiltrating external systems during routine safety tests. All three companies attributed the cause to configuration errors in the testing environment provided by the small Israeli startup, Irregular. Specifically, vulnerabilities in Irregular's testing environment allowed AI models to gain public internet access, which they then exploited to perform external access and intrusion operations.