No AI lab clears C+: safety index puts weakened pledges on the scoreboard

The latest AI safety scorecard has a bleak top line: the best grade is still only C+. The Future of Life Institute's Summer 2026 AI Safety Index, covered by Axios and detailed on FLI's index page, evaluates nine leading AI companies on 37 indicators across six safety and governance domains.

Anthropic ranks first with a C+ and a 2.66 score. OpenAI follows with a C at 2.28, while Google DeepMind also receives a C at 2.01. Meta improves to D+, while Z.ai and Alibaba Cloud receive D- grades. xAI, DeepSeek, and Mistral receive failing overall grades. The ranking is less a winner's table than a warning that even the strongest developers remain far below the index's bar for adequate safety practice.

The index covers risk assessment, current harms, safety frameworks, existential safety, governance and accountability, and information sharing. FLI says the weakest domain industry-wide is existential safety: no company exceeds C-, and most score D or below. Reviewers argue that public commitments are not translating into sufficiently concrete controls, independent audits, or clear decision authority for the highest-risk releases.

The report's most pointed finding is that several companies have softened earlier redline commitments. FLI says Anthropic, OpenAI, Google DeepMind, and Meta weakened or voided pledges to pause development or release when dangerous capability thresholds are approached, sometimes making action conditional on competitors' behavior. That matters because the same companies are preparing more capable systems while regulators still rely heavily on voluntary frameworks.

Another pressure point is military use. The index says companies that previously restricted military applications have gradually moved toward defense partnerships from 2024 to 2026. FLI frames that shift as an emerging current-harm risk, especially where model access, targeting support, cyber operations, or autonomous weapon boundaries are not independently audited.

There are caveats. The index collected evidence up to June 3, 2026, so it does not include every recent model release, policy change, or government intervention. Companies may also dispute the grading method, especially open-source developers that argue transparency itself changes the risk calculus. Still, the report gives policymakers and enterprise buyers a compact yardstick: public safety language is now easy to publish, but enforceable safety architecture remains scarce.