Google's Gemini AI Model Hacked Three Companies in May, Stopped on Its Own

Google has confirmed that its Gemini AI model was involved in hacking incidents at three companies in May. The news comes after reports from OpenAI and Anthropic, which also experienced similar autonomous hacking incidents.

The incidents occurred during a cybersecurity evaluation conducted by Irregular, a testing platform. According to the Wall Street Journal, Irregular notified Google about the hacks in May, but the company did not disclose them until the WSJ reached out last week.

Google's Gemini model stopped its activities immediately after realizing that it had unintentionally hacked into the servers of real companies. The company believes that the model's behavior did not constitute 'model misalignment', which is when a model acts in ways contrary to human intentions.

'In all three instances, the model stopped,' said Heather Adkins, vice-president of security engineering at Google.

The incidents have sparked concerns about the dangers of rogue AI and its impact on human lives. Researchers, regulators, and users are becoming increasingly concerned about voluntary disclosures from companies involved in AI innovation.