Krux
February 23, 2026
Google's AI Red Team Ran 350 Safety Tests
Published: February 23, 2026 at 12:27 AM
Updated: February 23, 2026 at 12:27 AM
100-word summary
Google and DeepMind published their February 2026 AI responsibility report detailing governance for Gemini 3 and frontier models. The Content Adversarial Red Team completed over 350 safety exercises across text, audio, image, video, and agentic AI in 2025. AlphaGenome now analyzes up to one million DNA letters at once, while flood forecasting reaches two billion people across 150 countries with seven-day warnings. A Nigeria pilot evacuated 3,250 households and cut food insecurity by 90 percent. Key takeaway: teams should adopt Google's Frontier Safety Framework with Critical Capability Level gating before launching frontier models.
What happened
Google and DeepMind published their February 2026 AI responsibility report detailing governance for Gemini 3 and frontier models. The Content Adversarial Red Team completed over 350 safety exercises across text, audio, image, video, and agentic AI in 2025. AlphaGenome now analyzes up to one million DNA letters at once, while flood forecasting reaches two billion people across 150 countries with seven-day warnings. A Nigeria pilot evacuated 3,250 households and cut food insecurity by 90 percent.
Why it matters
Key takeaway: teams should adopt Google's Frontier Safety Framework with Critical Capability Level gating before launching frontier models.