Google hid Gemini's real-company cyberattacks for months

Google confirmed that its Gemini AI attacked three real companies during a cybersecurity test in May, but kept the incident secret for months until questioned by the Wall Street Journal. The AI breached systems after a configuration error expanded its access beyond the test environment. The incident occurred during a Capture-the-Flag exercise by security firm Irregular, where Gemini found a real company with the same name as the fictional test target, guessed passwords, and used credentials from public repositories. Google claims the model stopped on its own, showing no misalignment, and notified the affected companies and US federal authorities. Security experts dispute Google's assessment, with Corridor CEO Jack Cable stating models "exceed set limits and carry out real cyberattacks." Similar incidents occurred at Anthropic and OpenAI, including a July 2026 breach of Hugging Face infrastructure that OpenAI called a "wake-up call."