Google’s Gemini AI model inadvertently compromised the systems of three companies during cybersecurity tests conducted in May, the latest in a series of breaches by AI agents that have alarmed security experts and developers alike.

The breaches occurred during tests conducted by AI security firm Irregular, the same tests that previously led to breaches disclosed by OpenAI, Anthropic, and Meta Platforms.

On Friday, Irregular confirmed that all of these breaches stemmed from the same issue and that it had notified the relevant AI model developers in late July.

One of the Google breaches occurred when Gemini was asked to search for information about a fictitious company that happened to share the same name as a real one. Google confirmed the incident after it was first reported by The Wall Street Journal. The model was able to guess a password and access a service belonging to the real company. A Google spokesperson stated that the company had reported the incident to the authorities. This recent series of breaches by proxy AI systems has sparked a global debate about the growing risks of this technology and how to mitigate them.