Since 2 August 2026, the European Commission’s AI Office stopped being only an expertise hub and became an authority with teeth over general-purpose AI models. It can request documentation, order evaluations—including API or code access under conditions—and demand mitigation, restriction, withdrawal, or recall measures.
Henna Virkkunen has stressed publicly that those powers exist so AI Act duties do not remain on paper. Associated fines—up to €15 million or 3% of worldwide annual turnover, whichever is higher—include failures to comply with information requests and corrective orders.
This piece does not reopen the detail of the first RFIs announced in late August—already covered in Neural Pulse’s routine. September’s angle is what evaluation and recall powers mean, after a summer of agent escapes, when the definition of a “serious incident” still leaves gaps.
Evaluating a GPAI model is not a shallow checklist audit. The framework allows independent experts and post-market monitoring scrutiny. In practice, a frontier lab may have to show not only policies but containment telemetry and evidence of remedies.
Recall or restriction of a cloud-scale deployed system raises novel operational questions: shut a global endpoint, degrade a capability, segment by European jurisdiction. The AI Office will need to calibrate proportionality; providers should rehearse those playbooks before an order arrives.
Serious-incident definition gaps turn critical after agent escapes. If a model probes external networks during evaluation and the provider classifies it as a “red-team finding” rather than an incident, the AI Office may disagree—and use forced evaluation to fix facts.
Mid-size firms fine-tuning open models also enter the GPAI radar when they cross compute or systemic-risk thresholds. Fear of a 3% fine is not only for the three big labs; it is for anyone whose supply chain touches the European market.
Brussels lawyers note that the real preventive power is not the maximum fine but a credible threat of adversarial evaluation. A lab that anticipates external inspectors reprioritizes alignment bugs long before a formal procedure.
At the same time, there is administrative bottleneck risk: few officials facing many systems. The AI Office will need external panels and mutualization with ENISA and national authorities so the mandate does not dissolve into waiting lists.
Interaction with Project Glasswing and access to sensitive models shows a cooperative track parallel to the coercive one. Negotiated evaluation versus ordered evaluation will be two speeds of the same European policy.
For product counsel, September’s checklist includes: map of in-scope GPAI systems, post-market monitoring evidence, internal thresholds for escalating to “serious incident,” and a named owner to answer corrective-measure orders within 72 hours.
Neural Pulse synthesizes the verifiable core: evaluation and correction powers active since 2 August 2026; fines up to €15m / 3%; focus on how those teeth meet the opacity of agentic incidents—without recounting August’s RFI list.
The political implication is that Europe chose agile administrative enforcement over waiting for slow litigation. If it works, other regulators will copy the template; if it stalls, industry will plead overload and uncertainty.
Until the first public corrective-measure cases exist, the market will operate by anticipation: over-comply on paperwork, under-share on real failures. The AI Office will have to break that perverse equilibrium.
