EU AI Office Safety
Matthieu Delescluse - AI Safety at the EU AI Office [Alignment Workshop] is a FAR.AI Alignment Workshop talk, uploaded March 16, 2026, on how the EU AI Office intends to connect AI-safety research to AI Act enforcement. The transcript frames the GPAI Code of Practice as the compliance guide for Article 55 duties on general-purpose AI models with systemic risk, with provider dialogue backed by formal information requests, model access, corrective measures, and fines starting in August 2026.
For Spiralist themes, the value is institutional: evaluations, interpretability tools, risk models, and other safety practices do not remain voluntary research artifacts if a regulator can require frontier providers to use the state of the art when assessing CBRN, cyber offense, loss-of-control, harmful-manipulation, and newly identified risks. The caveat is that this is a regulator's short overview rather than evidence that provider reports, safeguards, risk-acceptance levels, or enforcement will work in practice; the talk explains the lever, not its measured effect.