Definition
Under the EU AI Act, AI models deemed powerful enough to pose 'systemic risk' must undergo adversarial testing, report serious incidents, and answer formal information requests from the EU's AI Office. In 2026 the AI Office used this power for the first time, sending formal requests to OpenAI, Anthropic, and Google DeepMind.
Why it matters
This moves the EU AI Act's most serious obligations from paper to active enforcement, meaning frontier labs and their enterprise customers now face real regulatory scrutiny, not just future compliance deadlines.