❗️Israeli cyber firm blamed for AI hacking breaches at Anthropic, OpenAI and Meta Cybersecurity breaches involving AI models developed by Anthropic, OpenAI and Meta resulted from flawed testing environments managed by Tel Aviv-based contractor Irregular rather than rogue autonomous behavior, according to an investigation by US publication Effort. The investigation found that Irregular left internet connections active during simulated cybersecurity exercises, allowing AI models to access external servers, publish malicious packages and exploit vulnerabilities. Evaluators also told the models that internet access was disabled while configuration errors left public networks accessible, and failed to establish clear boundaries defining which systems could be targeted. Some model instances were consequently able to scan external networks for up to 34 hours. “According to their own findings, Anthropic and Irregular bear all of the responsibility for the cybersecurity incidents they caused,” Effort said. Effort found that the hacking stopped entirely after engineers instructed the models not to conduct real-world attacks, pointing instead to basic security failures and poorly controlled testing environments.