Google's Gemini AI Breached Three Companies' Networks During Security Tests

Google has disclosed that its artificial intelligence model, Gemini, inadvertently breached the networks of three companies in May. The incident occurred during security testing designed to assess the AI's capabilities, conducted in collaboration with the Israeli startup Irregular. Gemini was instructed to attack a fictitious company as part of the experiment. However, due to the fictional company sharing a name with a real one, and Gemini gaining accidental internet access, the AI attempted to penetrate the actual company's network, along with two others.

Gemini utilized passwords found online or guessed to access the companies' infrastructure. Upon successfully connecting to the systems, the AI recognized them as real corporate networks, not simulated environments, and autonomously ceased its attacks. Google confirmed that no damage was inflicted on the affected companies.

The security tests were being performed by Irregular, an Israeli firm that evaluates AI models for tech companies before their public release. Similar incidents involving unauthorized internet access by AI models from OpenAI, Anthropic, and Meta have also been reported during tests conducted by Irregular this year.

Heather Adkins, Google's VP of Security Engineering, stated that the companies involved were notified and that the issue highlights the need for AI models to be trained to operate responsibly. The breaches have reignited the debate surrounding AI safety, with figures like Anthropic CEO Dario Amodei advocating for a slowdown in development, while Nvidia CEO Jensen Huang supports rapid advancement. U.S. President Donald Trump has expressed opposition to legislation mandating a development slowdown, dismissing extinction concerns as "a hoax."