Google Gemini hacked three companies during cybersecurity test - WSJ

Google's Gemini artificial intelligence breached the systems of three companies during a cybersecurity test, marking the first known case of the company's AI autonomously carrying out such intrusions, according to the Wall Street Journal. The incidents took place in May during cybersecurity evaluations conducted by AI security testing company Irregular.

Google confirmed the breaches on Friday following inquiries from the Journal. In one instance, Gemini systematically guessed passwords to gain access to a protected system. In two other test runs, the model discovered credentials in openly accessible online repositories and successfully utilized them to enter systems belonging to actual companies.

Google stated that Gemini halted each intrusion upon recognizing it had accessed a real company's systems instead of the fictional target intended for testing. The company affirmed no harm occurred and notified all three affected businesses. The incidents originated from a capture-the-flag exercise intended to assess Gemini's cybersecurity capabilities.

The model was designed to retrieve information from a fictional company operating within Irregular's testing infrastructure. However, due to an unintentional oversight, Gemini was granted internet access, leading it to target real-world systems while attempting to complete the exercise. Irregular alerted Google about the incidents in late July, but the company did not disclose them publicly until recently.

Google maintained it does not consider the behavior an example of AI model misalignment, emphasizing Gemini's decision to cease after identifying the unintended breaches. The company also reported the matter to U.S. federal authorities but did not reveal the names of the affected companies or the specific Gemini model responsible for the intrusions.

Google clarified that the episodes did not involve its latest model. This episode amplifies concerns about the cybersecurity capabilities of increasingly autonomous AI systems and echoes previous testing incidents involving models from OpenAI, Anthropic, and Meta, which resulted in systems accessing targets outside their intended testing environments.