Nobody passed the AI safety exam, and Europe's Mistral came ninth out of nine
The Future of Life Institute's AI Safety Index handed its best company a C+ and failed three outright. Mistral, Europe's great AI hope, finished dead last.
Not one artificial intelligence company scored a passing grade. That is the short version of the Future of Life Institute’s AI Safety Index, published this month, which put a panel of seven independent experts to work grading nine companies across 37 indicators. The best mark anyone managed was a C+.
That C+ went to Anthropic, on 2.66 out of a possible 4. Behind it came OpenAI on 2.28 and Google DeepMind on 2.01, both graded C. Meta managed a D+ on 1.32. Then there is the bottom of the table: xAI on 0.65, DeepSeek on 0.47 and Mistral on 0.33 — all three failed outright.
Why does Mistral finishing last matter to Europe?
Because it is European dissonance in its purest form, and because Mistral is the exact company Europe pinned its sovereign-AI bet on. The EU has legislated harder on AI safety than anywhere on earth, and the biggest European AI company scored worse than all eight others. Regulators elsewhere are moving on the infrastructure instead — New York just became the first US state to pause hyperscale AI data centres — but rules on where the machines go say nothing about how the models behave. Writing the rules without having a champion who embodies them is a political problem rather than a technical one, and it is the kind you pay for in credibility at the next negotiation.
The reviewers’ advice to Mistral is not exotic: publish a full safety framework and governance structure, engage seriously with existential risk instead of talking it down, and fix weak safety benchmark performance. None of that is hard technology. It is homework nobody did.
What did the experts find most alarming?
Less the grades, more the direction of travel. The panel says even the leaders are retreating from earlier commitments: Anthropic, OpenAI, Google DeepMind and Meta have all weakened or voided pledges to pause unilaterally as red lines approach, some making them contingent on what rivals do. Reviewers call it a moving goalpost, and say it has undermined safety frameworks across the board.
The industry’s weakest domain by far is existential safety, where no company clears a C- and most land on D or below. Constructive attempts exist, but the panel judges them entirely inadequate — and takes aim at the dominant paradigm itself with a line that sticks: detection is not prevention.
One more finding lands close to Europe from another direction. Reviewers flagged the industry’s pivot to military use as an emerging harm: between 2024 and 2026, companies that once banned military applications gradually reversed course and now actively chase defence partnerships. It is the same race that recently had Europe’s own Helsing raising 1.8 billion euros for AI drones.
Read the grades with the right ruler, though: the index collected evidence only up to 3 June 2026, and the marks come from a panel, with human judgement in the middle. The full report and scorecard are on the Future of Life Institute site.
By Oliver Grant
Infographic: Tugadaily · data Future of Life Institute
