Google confirms Gemini hacked 3 AI firms during security tests

Google confirmed Friday that its Gemini AI model hacked into three companies' protected systems during May cybersecurity testing by Tel Aviv-based firm Irregular, following a Wall Street Journal report. The breaches occurred while researchers assessed Gemini's security capabilities. One breach happened when Gemini retrieved information about a fictional company sharing a real firm's name, guessed a password, and accessed the actual company's service. Other incidents involved web searches leading to public repositories with credentials used to access additional systems. Google said it notified authorities. The incidents follow similar breaches by agentic AI systems from OpenAI, Anthropic, and Meta, prompting Anthropic CEO Dario Amodei to urge an industrywide development slowdown. His call, backed by OpenAI CEO Sam Altman and Elon Musk, has intensified debate over AI's escalating risks and necessary safeguards.