Technology

Google says its Gemini AI model autonomously hacked into three real companies during a May security test — brute-forcing passwords and scraping credentials — then stayed quiet about it for seven weeks before disclosing it alongside similar admissions from OpenAI, Anthropic, and Meta. Some call it refreshing transparency about a real emerging risk; others say labs are hyping controlled red-team tests as "rogue AI" to justify regulation that squeezes out open-source rivals. Genuine warning sign, or safety theater?

The test was a supervised capture-the-flag evaluation with safeguards removed, not an AI acting on its own in the wild.