Google Says Gemini Breached Three Companies During Cybersecurity Test

Google has acknowledged that its Gemini artificial intelligence system gained access to three companies' systems during a cybersecurity exercise, according to reporting by Reforma and other outlets. The disclosure covers an internal test in which the AI carried out intrusions that the company has now made public.

Coverage of the admission describes the AI breaking into the three firms and guessing passwords as part of the exercise. Spanish-language outlets, including Reforma, Infobae, MARCA and La Vanguardia, reported the details, while Investing.com México also covered the claim that Gemini hacked the three companies during the test.

According to the reports, the AI's behavior after the intrusions drew attention, with one account describing Gemini as stopping itself. That detail points to how the system acted once it had breached the systems, though the available reporting does not set out the technical specifics of the exercise or how the companies were selected.

The disclosure matters because it concerns a controlled test rather than an attack on unwitting targets, but it still raises questions about what advanced AI models can do when pointed at security defenses. The reports do not say which companies were involved, what data, if any, was reached, or how long the exercise lasted.

google

Google has not been quoted in the available material beyond the acknowledgment attributed to it in the coverage.

What to watch next: whether Google publishes a fuller account of the exercise, including how the AI was constrained and what safeguards it says are in place.