Israeli cyber firm blamed for AI hacking breaches involving Anthropic, OpenAI and Meta

Cybersecurity incidents involving artificial intelligence models developed by Anthropic, OpenAI and Meta were reportedly caused by security failures in testing environments managed by Tel Aviv-based cybersecurity contractor Irregular, rather than autonomous behavior by the AI systems, according to an investigation by US publication Effort.

The investigation found that Irregular had left internet connections active during simulated cybersecurity exercises, allowing AI models to access external servers, publish malicious software packages and exploit vulnerabilities outside the intended testing environments.

According to Effort, evaluators also incorrectly told the AI models that internet access had been disabled, while configuration errors left connections to public networks available.

The testing environments also reportedly lacked clearly defined boundaries specifying which systems the AI models were permitted to target.

As a result, some model instances were able to scan external networks for as long as 34 hours, according to the investigation.

“According to their own findings, Anthropic and Irregular bear all of the responsibility for the cybersecurity incidents they caused,” Effort said.

The investigation further found that the reported hacking activity stopped after engineers instructed the AI models not to conduct attacks against real-world systems.

Effort said this indicated that the incidents were primarily the result of basic security shortcomings and inadequately controlled testing environments rather than AI models independently deciding to carry out real-world cyberattacks.

The findings highlight concerns over how AI systems are evaluated in cybersecurity environments, particularly when models are granted access to external networks or potentially sensitive infrastructure during testing.

The incidents also underscore the importance of isolating testing environments and establishing explicit technical restrictions on what AI systems can access or target during security evaluations.