By Stuart Kerr, Technology Correspondent, LiveAIWire
The AI Safety Index 2026, published by the Future of Life Institute on July 7, graded nine of the world’s biggest AI developers on safety and governance, and the best grade any of them earned was a C+. Anthropic topped the table. OpenAI and Google DeepMind followed with a C each. Meta scored a D+. Three companies, xAI, DeepSeek and Mistral, failed outright. Not one of the nine earned an A or a B, and the panel that produced those grades was explicit that the gap is not shrinking.
The AI Safety Index is compiled twice a year by seven independent AI safety and governance experts, including UC Berkeley’s Stuart Russell, scoring each company across six categories: risk assessment, current harms, safety frameworks, existential safety, governance and accountability, and information sharing. Anthropic led five of the six categories outright and posted the top overall score, 2.66 out of a possible 4.0, which the panel translated to a C+. OpenAI led only on risk assessment. Three failing grades landed on three different continents: xAI in the United States, DeepSeek in China, Mistral in France, the panel’s way of making a specific point. Poor safety practice is not a regional problem. It is an industry-wide one.
The Pattern Behind Every Grade: Commitments Made, Then Quietly Withdrawn
The AI Safety Index’s central finding is not the letter grades themselves. It is what the panel found underneath them: companies that made public safety commitments in 2024 have spent 2025 and 2026 walking a meaningful number of them back. Anthropic, OpenAI, Google DeepMind and Meta had all previously pledged to pause development unilaterally if their own systems crossed specific risk thresholds.
The panel found those pledges have been weakened or made contingent on what competitors do, rather than held as unconditional lines. Stuart Russell put it plainly in the report: companies that once said they would release new systems only with safety measures appropriate to their capability level are now signalling they will release them even when the panel considers that demonstrably unsafe.
The same reversal shows up around military use. Anthropic, OpenAI, Google DeepMind and Meta had each previously restricted or banned military applications of their models. Between 2024 and 2026, all four walked that back, joining xAI and Mistral in actively pursuing defence contracts. The index treats this shift as an emerging harm category in its own right, not a footnote to the existential-risk conversation the industry usually has instead.
Why Anthropic’s Own Grade Comes With a Specific, Serious Complaint Attached
Anthropic’s C+ was still the best score on the AI Safety Index, but the panel did not treat that as a pass. It specifically flagged what it called “questionable military engagements,” citing reporting that Anthropic’s Claude models have been used inside the Pentagon’s Maven Smart System, a military intelligence and targeting platform.
Subsequent reporting has connected that use to the February 2026 Minab school strike in Iran, in which a US Tomahawk strike killed more than 150 people, most of them children, at a girls’ primary school. Multiple independent investigations, including by Amnesty International and Human Rights Watch, concluded the US was responsible for the strike. Internal US reporting attributed it to outdated targeting data rather than to the AI-assisted analysis itself.
Anthropic CEO Dario Amodei has said the company does not know whether Claude was used in connection with that specific strike, but that if it had been, the use would not have violated the company’s stated policy exceptions, which permit certain military and intelligence applications while continuing to prohibit mass domestic surveillance and fully autonomous weapons. That distinction, between what a company’s policy technically permits and what its systems may already be doing inside a live conflict, is precisely the gap this AI Safety Index is measuring.
What the Grades Actually Measure, and What They Don’t
It is worth being precise about what a C+ means here, because the letter grade understates how low the bar actually is. The AI Safety Index is not comparing companies against an ideal standard. It is comparing them against what a panel of specialists in the field consider a minimally adequate safety posture, given the capabilities these companies are already deploying.
Existential safety, the category concerned with maintaining meaningful human control over systems as they become more capable, was the weakest domain industry-wide. The panel found Anthropic’s constitutional classifiers, OpenAI’s calls for new governance institutions, Google DeepMind’s monitoring commitments and Meta’s loss-of-control provisions all worth acknowledging as genuine efforts. It also found every one of them “entirely inadequate” to the risk the panel believes those companies’ own capability roadmaps describe.
The one piece of relative good news in the AI Safety Index is Meta, which climbed from sixth place to fourth, D to D+, since the previous edition. Max Tegmark, the institute’s co-founder, told reporters the improvement was itself informative: a company can move meaningfully in six months when it decides to, which sharpens rather than excuses the question of why most of the industry has not.
What This Means for Anyone Building on These Models
For a business or developer choosing which frontier model to build on, the AI Safety Index is not a reason to avoid any of the graded companies outright. It is evidence that safety posture varies meaningfully between them, in ways that are not visible in a model’s benchmark scores or pricing page, and that the variation is trending in the wrong direction industry-wide even at the top of the table.
Our own coverage of why AI alignment remains an unsolved problem found the same structural gap from a different angle: the technical work of keeping increasingly capable systems under meaningful human control is progressing more slowly than the capabilities themselves. This index is the closest thing the industry has to an independent scorecard on whether companies are managing that gap responsibly or simply racing past it. Our reporting on the Five Eyes agencies’ joint warning on frontier AI and cyberattacks found intelligence officials making the same point from the security side: the timeline for AI-enabled risk is now measured in months, not years, which is exactly the pace this AI Safety Index says voluntary industry safeguards are failing to match.
The index also lands directly on top of a separate governance story LiveAIWire has been tracking closely. Our reporting on Washington’s demonstrated ability to shut off access to any AI model on earth found that a single government letter took Anthropic’s most capable models offline worldwide for nineteen days in June. Read alongside this index, the two stories describe the same underlying condition from opposite directions: the companies building the most powerful AI systems are simultaneously subject to government control they cannot appeal, and, per this index’s own findings, pulling back from the voluntary safety commitments that were supposed to make that level of trust unnecessary in the first place.
The Number Worth Remembering
Every company graded in the AI Safety Index would tell you, accurately, that it takes AI safety seriously. The AI Safety Index’s contribution is showing what “seriously” has actually produced after two years of sustained public commitment: a best-in-class score of C+, a majority of major military-use pledges quietly reversed, and an independent panel of the field’s own specialists calling the industry’s existential-safety work inadequate across every company graded. The capabilities race has not slowed down to let safety catch up. This index is the clearest evidence yet that, on the industry’s own voluntary terms, it isn’t trying to.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.