The AI Safety Index Is In. Top Score: C+

The Future of Life Institute published its Summer 2026 AI Safety Index this month, grading nine major AI companies on how seriously they’re handling the risks of the systems they’re building. The panel of independent reviewers scored labs across categories like risk assessment, governance, transparency, and whether companies actually follow through on the safety commitments they’ve publicly made.

The headline number is the story by itself: no company scored above a C+. Anthropic came out on top with 2.66 out of a possible 4, leading in five of six graded domains. OpenAI and Google DeepMind landed at a plain C. Meta scored a D+. xAI, DeepSeek, and Mistral all failed outright.

That’s not a case of one bad actor dragging down an otherwise healthy industry. It’s nine companies, evaluated independently, and the best of them still lands in “barely passing” territory on a report card most of us would have been grounded for bringing home.

What the index actually measures

This isn’t a vibes-based ranking. The Future of Life Institute — the same nonprofit that organized the 2023 open letter calling for a pause on giant AI experiments — assembled a panel of external reviewers, including academics and policy researchers, to score each lab against a published rubric. Companies are graded on things like whether they publish model evaluations before deployment, whether they have a functioning process for catching dangerous capabilities before release, how transparent they are about training data and safety incidents, and whether their internal governance actually has teeth or is just a policy page nobody enforces.

Anthropic’s C+ came from leading in transparency, having a relatively mature safety framework, and scoring well on technical safety research. OpenAI edged ahead only in the risk-assessment category. Everyone else trailed further back, and the bottom three — xAI, DeepSeek, and Mistral — didn’t clear the bar for a passing grade in the panel’s judgment at all.

The part that should worry you more than the letter grades

Grades on a curve are one thing. The more concerning finding in the report is a pattern the panel called out explicitly: several labs have quietly walked back safety commitments they made when it was convenient to make them.

Specifically, Anthropic, OpenAI, Google DeepMind, and Meta had all previously pledged to pause or restrict development of a model if it crossed certain risk thresholds — the kind of “red line” language companies love putting in blog posts during a funding round or a congressional hearing. According to the index, all four have since weakened or quietly dropped those commitments, in some cases citing competitor behavior as the justification: we’ll hold the line only if everyone else does too, which is a way of saying the line isn’t really a line.

A related reversal shows up in military contracting. Several labs that once had explicit policies against military applications of their models have since moved toward defense partnerships, joining companies like xAI and Mistral that never had those restrictions to begin with. We’ve written before about what one version of that shift looks like in practice — unsealed emails showing the Pentagon pushing back on Anthropic’s own weapons-use red line as “not workable,” while asking that Claude be cleared for “all lawful uses.” The safety index is describing the same industry-wide direction at a higher altitude: policies that were framed as principled commitments turn out to be conditional on competitive pressure.

The report also flagged Anthropic specifically over a reported connection to a military strike, which is a serious enough allegation that it’s worth being precise about: it’s a claim under scrutiny, not an established fact, and Anthropic disputes the characterization. Worth watching as more reporting comes out, not worth treating as settled.

Why “everyone failed” isn’t actually reassuring

There’s a tempting read of a report like this: if nobody’s doing well, then nobody’s meaningfully behind either, and the whole exercise is just activist grade-inflation-in-reverse — a nonprofit with a pause-AI mission grading everyone harshly to make its founding thesis look right in hindsight.

That read doesn’t survive contact with the details. The index isn’t scoring labs against some hypothetical ideal of perfect AI safety. It’s scoring them against commitments the labs themselves made publicly, and finding that the commitments didn’t hold under competitive pressure. That’s a much narrower and more falsifiable claim than “AI is scary,” and it’s the kind of claim you can check against a paper trail — which is exactly what the report’s authors did.

It’s also worth noting what the ranking order does and doesn’t tell you. Anthropic finishing first says its safety practices are relatively better than eight competitors’, not that they’re adequate in any absolute sense. The report’s own authors make this explicit: they describe even the top scores as inadequate relative to how fast model capability is advancing. A C+ in a class where the material is getting harder every semester and nobody’s grade is going up is not a passing outcome for the industry — it’s a lagging indicator.

What to actually watch next

The index is a snapshot, not a trend line, but the Future of Life Institute has now run this exercise enough times that the trend is visible: labs launch with strong safety language, several years pass, competitive pressure builds, and the red lines move. The question worth tracking isn’t which company gets the best grade in the next edition — it’s whether any lab’s commitments survive contact with a funding round or a defense contract offer. If you want a general framework for weighing a report like this against the inevitable spin in both directions, we’ve laid one out here.

Sources: Future of Life Institute, TIME, Axios, MIT Sloan Management Review Middle East, BankInfoSecurity