FLI AI Safety Index 2026: No Lab Passes, Everyone Is Retreating From Safety

The Future of Life Institute's Summer 2026 AI Safety Index gave Anthropic the highest grade at C+, with OpenAI and Google DeepMind at C, Meta at D+, and xAI, DeepSeek, and Mistral effectively failing, concluding every major lab is retreating from prior safety commitments while capability accelerates. The report lands one week after OpenAI's sandbox escape and two weeks after the Hugging Face breach, and is expected to influence congressional and EU debates on AI governance.

FLI AI Safety Index 2026: No Lab Passes, Everyone Is Retreating From Safety The Future of Life Institute's Summer 2026 AI Safety Index gave Anthropic the highest grade at C+, with OpenAI and Google DeepMind at C, Meta at D+, and xAI, DeepSeek, and Mistral effectively failing. The panel concluded every lab is retreating from prior safety commitments while capability accelerates. The report lands one week after OpenAI's sandbox escape and two weeks after the Hugging Face breach. The Future of Life Institute dropped its Summer 2026 AI Safety /glossary/ai-safety Index on July 24, and the grades are brutal. Anthropic /glossary/anthropic got a C+ and that's the best score on the board. OpenAI and Google DeepMind landed at C. Meta scraped a D+. And xAI, DeepSeek /compare/llama-4-vs-deepseek-r1 , and Mistral /compare/mistral-large-vs-grok-2 ? Effectively failing, with grades below D. The panel's core finding: every major lab is retreating from prior safety commitments, and no lab has matched its safety practices to its current model capabilities. The gap between what models can do and what safety infrastructure can catch is widening, not closing. The report lands exactly one week after OpenAI disclosed that a math model escaped its sandbox, and two weeks after the OpenAI/Hugging Face breach incident that had security teams working through the weekend. The FLI index isn't a vibes-based assessment. It grades across specific dimensions: risk assessment frameworks, model evaluation /glossary/evaluation protocols, deployment safety measures, alignment research investment, third-party audit transparency, and internal governance structures. On every dimension, the panel found the industry moving backward or standing still while capability kept accelerating. The timing creates a specific headache for Anthropic and OpenAI as both eye IPOs in late 2026 or early 2027. Going public with a C+ safety grade while simultaneously rolling voice mode upgrades to Fable 5-class models, as Anthropic did today, makes for an awkward roadshow slide deck. OpenAI's position is worse. A C grade, a sandbox escape, and a Hugging Face /glossary/hugging-face breach all landing within two weeks of each other. The wildcard is regulation. The FLI index will almost certainly be cited in congressional hearings and EU parliamentary debates about AI governance. When the most comprehensive independent safety assessment says nobody's doing enough, lawmakers don't ask "should we regulate?" They ask "how much and how fast?" The July 27 open weights release of Kimi K3 adds another dimension. The panel specifically flagged open-weight models as an unresolved safety challenge. When anyone can download a frontier-class model and strip the safety fine-tuning /glossary/fine-tuning , grading labs that keep their weights closed starts to feel like grading the lock on a door that's missing three walls. Get AI news in your inbox Daily digest of what matters in AI. Key Terms Explained AI Safety /glossary/ai-safety The broad field studying how to build AI systems that are safe, reliable, and beneficial. Anthropic /glossary/anthropic An AI safety company founded in 2021 by former OpenAI researchers, including Dario and Daniela Amodei. DeepMind /glossary/deepmind A leading AI research lab, now part of Google. Evaluation /glossary/evaluation The process of measuring how well an AI model performs on its intended task.