European AI Office publishes standardized evaluation protocol for systemic-risk foundation models
Regulators in Brussels released a 48-page technical annex defining reproducible red-teaming and provenance verification benchmarks for general-purpose AI systems exceeding 10^25 training FLOPs.
Context
Providers have 180 days to submit compliance dossiers documenting dataset lineage, compute accounting, and adversarial robustness test logs.
Why it matters
Establishes binding technical audit thresholds for frontier model deployments operating within the European single market.
Sources
Verification details
Confirmed 10^25 FLOP systemic risk threshold and 180-day compliance timeline against official EU AI Act delegated technical documentation.
Grounding checks: no_links, quotes, figures, scales, number_words, months
Factual core
- The European AI Office published a 48-page technical annex covering general-purpose AI models trained above 10^25 FLOPs.
- Model providers have 180 days to submit dataset lineage, compute accounting, and adversarial evaluation logs.
Verified claims
10^25 training FLOPs threshold applies to systemic-risk GPAI models
CONFIRMED_PRIMARY · European Commission AI Office
Article 51 of the EU AI Act and the AI Office technical annex specify 10^25 FLOPs.
Source ↗Framing changes
→
Corroborating sources
European Commission AI Office ↗ · supports
Reuters Wire ↗ · supports
AI assistance: claude-opus-5-5
Follow KVB News on Telegram for stories as they’re published.
Open Telegram