AI Safety Index 2026: No Company Scored Above a C+

The Future of Life Institute’s Summer 2026 AI Safety Index just graded the nine biggest AI labs on safety, and not one of them cleared a C+. Anthropic came out on top with a 2.66 (still just a C+), OpenAI and Google DeepMind landed in the C range, and three labs — xAI, DeepSeek, and Mistral — flat-out failed. I’ve been tracking this report since its first edition, and the pattern this time is a little unsettling: the labs racing hardest to ship new models are also the ones quietly loosening the safety promises they made a year or two ago.

I dug through the full report and the coverage around it to pull out what actually matters for anyone following where AI regulation and corporate accountability are headed.

What the AI Safety Index actually measures

The Index is put together twice a year by the Future of Life Institute (the nonprofit behind the 2023 “pause giant AI experiments” open letter), using an independent panel of seven experts: David Krueger, Sharon Li, Tegan Maharaj, Sneha Revanur, Stuart Russell, Robert Trager, and Yi Zeng. That’s a mix of machine learning researchers, governance specialists, and AI policy advocates — not company insiders grading their own homework.

This summer’s edition, published July 7, 2026, scored nine companies across 37 indicators grouped into six domains: risk assessment, current harms, safety frameworks, existential safety, governance and accountability, and information sharing. Grades ranged from C+ down to F, using public disclosures and material the companies submitted themselves.

The grades, ranked

Anthropic held the top spot again, leading five of the six domains thanks to relatively strong transparency and a more developed safety framework than its rivals. But “leading” is doing a lot of work in that sentence — a 2.66 out of a possible 4.0 is still a middling grade, and the panel was blunt that even the frontrunner has real gaps.

Why the top score keeps stalling at C+

This is the part that stood out most to me. The report’s authors describe a “moving goalpost” problem: Anthropic, OpenAI, and Google DeepMind have all walked back or loosened earlier pause and safety-testing commitments they made when the AI safety conversation first went mainstream. Anthropic itself dropped a pledge — reported back in February 2026 — to avoid training a new system unless it could guarantee its safety measures were adequate in advance. The panel had specifically recommended reversing that call, and instead watched it happen.

UC Berkeley professor Stuart Russell, one of the seven panelists, put it plainly: “While there is good work being done on AI safety in the industry, the capabilities race has become more extreme.” That’s not a fringe opinion from an outside critic — it’s coming from inside the evaluation process itself.

The military pivot nobody saw coming a year ago

A second theme jumped out across multiple domains: companies that once publicly ruled out military applications for their models are now actively pursuing defense contracts. Anthropic, OpenAI, Google DeepMind, and Meta have all reversed earlier bans on military use cases in some form. The report treats this as an emerging-harm category on its own, separate from the usual jailbreak-and-misuse concerns, because it changes who the customer is and what “safe deployment” even means.

Meanwhile the domain that scored worst across the board, for every single company, was existential safety — nobody cleared a C-. The panel called current industry approaches to catastrophic-risk planning “entirely inadequate,” which is a stronger phrase than these reports usually reach for.

It’s worth noting the rankings weren’t frozen in place, either. Meta actually climbed to fourth place this edition, up from where it sat a year ago, while xAI slid down to seventh — a reminder that these scores move with real policy changes at each company, not just reputation. Z.ai and Alibaba Cloud, two labs new to the Index this round, entered in the D- range, which the panel attributed mostly to thin public disclosure rather than confirmed unsafe practices — they simply haven’t published enough for evaluators to score them any higher.

None of this is happening in isolation from the money involved, either. These are the same nine labs currently locked in the most capital-intensive technology race in history, and the report’s authors are explicit that the safety slippage they’re documenting tracks closely with how aggressively each company is competing on raw model capability.

This isn’t happening in a vacuum, either. It lines up with what came out of the UN’s Global Dialogue on AI Governance in Geneva earlier this month, where an independent scientific panel told governments outright that no technical guarantee exists that advanced AI avoids catastrophic harm. Two very different bodies, similar week, same uncomfortable conclusion.

Why this matters even if you don’t work in AI

I get it — a nonprofit’s report card on nine companies can feel abstract. But these grades increasingly drive the direction of actual regulation. State-level laws like Colorado’s AI Act already lean on independent risk assessments like this one when defining what “reasonable care” looks like for AI developers. When the industry’s own safety pledges quietly erode while public trust reports stay flat or improve, it creates exactly the gap regulators point to when they push for binding rules instead of voluntary commitments.

It’s also just useful as a consumer signal. If you’re choosing which AI tool to build a business process around, or which chatbot to trust with sensitive data, “who scored best on an independent safety audit” is a more grounded question than “who has the flashiest demo this month.”

Frequently asked questions

What is the Future of Life Institute’s AI Safety Index?

It’s a semiannual report card that grades major AI companies on safety practices across six domains — including risk assessment, existing harms, and existential-risk planning — using an independent panel of seven AI safety and policy experts. The Summer 2026 edition, released July 7, 2026, is the latest in the series.

Which AI company scored highest in 2026?

Anthropic, with a C+ grade (2.66 out of 4.0). It led five of the six graded domains, but the panel emphasized that even the top score reflects real, unresolved gaps rather than a passing grade in any strong sense.

Why did no company score above a C+?

The panel pointed to a combination of weakened safety pledges, a broad industry shift toward military-use partnerships, and near-universal failure on existential-safety planning — no company scored above a C- in that specific domain, which the report called “entirely inadequate.”

Does a low AI Safety Index score mean a company’s products are unsafe to use day-to-day?

Not directly — the grading focuses on governance, transparency, and catastrophic-risk planning rather than everyday product reliability. But it’s a reasonable proxy for how seriously a company treats long-term accountability, which matters if you’re relying on their tools for anything sensitive.