The AI Safety Report Card Is In. Nobody Passed — and Europe’s Champion Came Last.

Last week, gafam.ai told you what Europe is building — praising Mistral's open-weight, sovereignty-first strategy as the continent's most viable path. This week, an independent safety assessment placed that same company dead last in the world. Intellectual honesty requires us to hold both facts at once, and to work out what the contradiction actually means.

It is more revealing than either the praise or the failing grade alone.

What the Index Found

On July 7, the Future of Life Institute released its Summer 2026 AI Safety Index — a twice-yearly assessment in which a panel of seven independent researchers grades leading AI labs across six safety domains, from present-day harms to existential risk.

The evidence ran to June 3, 2026, and covered nine companies.
The grades are unforgiving. The highest overall grade was a C+, awarded to Anthropic, which led five of six domains. OpenAI and Google DeepMind each received a C. Meta improved to a D+. And three companies failed outright — one from each major region: xAI in the United States, DeepSeek in China, and Mistral in Europe, which placed dead last of all nine.
The headline finding is not any single grade. It is that the class valedictorian earned a C+.

As TIME put it, even the company that built its entire brand on safety is, by this independent panel's absolute standards, doing a mediocre job — at the exact moment these systems are being wired into cybersecurity, healthcare and autonomous agents.

The Finding That Should Worry Everyone

Beneath the grades sits a more troubling pattern. The panel found that Anthropic, OpenAI, Google DeepMind and Meta have all weakened or voided earlier pledges to pause development if their systems approached specified danger thresholds. The reviewers called this moving the goalposts. In February, Anthropic — the top-ranked lab — dropped its pledge never to train a system unless it could guarantee its safety measures were adequate.

The panel also flagged the industry's pivot to military AI as an emerging present-day harm. Labs that once banned military applications have reversed course to pursue defence partnerships — and the report criticised even Anthropic for questionable military engagements. Existential safety was the weakest domain across the entire industry: every one of the nine companies was rated inadequate at containing the most serious risks of the systems they are racing to build.

FLI chair Max Tegmark's summary was blunt: the companies, he said, are sprinting toward a cliff.

The European Dissonance — And Mistral's Serious Rebuttal

Now the part that matters most for gafam.ai, and it is genuinely two-sided.
On its face, the result is a European embarrassment. The European Union is the undisputed global leader in AI safety regulation — the EU AI Act is the most comprehensive framework in the world. Yet Europe's flagship AI company scored last on safety practices. Europe writes the rules; its champion, per this index, fails the test. That gap between regulating safety and demonstrating it is a real and uncomfortable finding.

But Mistral's response is not a dodge, and it deserves to be taken seriously. Mistral did not complete FLI's survey, and it disputes the methodology directly. Its argument, given to Axios: Mistral's models are open-weight, which means enterprises decide how they are fine-tuned and deployed and can build in the specific safety controls their context requires. And it turned the safety question back on the index itself — a handful of companies deciding, behind closed doors, what is safe for everyone else is a risk that Mistral would also highlight; open, independently scrutinised models are the check on that concentration of power.

This is a substantive philosophical dispute, not spin. FLI's framework is built largely around the closed-lab safety paradigm: pre-deployment testing, internal safety frameworks, unilateral pause commitments. Open-weight models have a fundamentally different safety architecture — distributed, deployer-controlled, and open to independent inspection by anyone. An index designed around closed-model assumptions will, almost by construction, rate the open-model approach poorly. Whether that reflects Mistral being genuinely less safe, or the framework measuring the wrong things for open models, is exactly the question the grade alone cannot answer.

The Honest Complication for Europe

gafam.ai will not resolve this in Europe's favour by reflex. Both things can be true: Mistral may indeed lag on the formal safety practices FLI measures, and FLI's framework may indeed undervalue the open-weight model of distributed safety. The failing grade is a real signal that Europe's champion has thinner formal safety infrastructure than the top American labs. The methodological critique is a real signal that the dominant definition of AI safety was written by and for closed labs.

What this exposes is a genuine European dilemma that the continent has not resolved. Europe's sovereignty strategy — the one we described approvingly last week — runs through open weights: Mistral, the EUROPA consortium, on-premise deployment. But open weights are precisely what the prevailing safety paradigm rates lowest, because once weights are released, no lab can pause, recall, or unilaterally control them.

So Europe's two instincts are in tension: its openness-and-sovereignty instinct, and the safety-first instinct its own regulation partly embodies. Is open-weight AI Europe's path to independence, or a safety liability it will be blamed for? Europe is currently pursuing both values without admitting they can conflict.

One more note on fairness: FLI is not a neutral referee. It is an advocacy organisation with a defined worldview — Tegmark drafted a 2025 letter calling for a ban on superintelligence, and the panel is composed of researchers with strong catastrophic-risk concerns. That does not invalidate the index; the transparency and pause-commitment findings are serious and well-documented. But the framework encodes a particular philosophy of what safety means, and that philosophy is itself part of what is being contested.

The European Perspective