Nobody passed, and everybody shipped
Nine frontier labs graded on safety; the best was a C+. In the same index, every lab that once banned military applications had quietly stopped banning them.
The Future of Life Institute graded nine frontier labs and none scored above C+. Buried in the same document: the labs that prohibited military applications have all reversed course.
Absolute standards are why nobody looks good
37 indicators, six domains, a panel of seven researchers and governance experts, graded against absolute performance standards rather than a curve. Anthropic led at 2.66 on a four-point scale; OpenAI and Google DeepMind followed at 2.28 and 2.01; Meta took a D+; Z.ai and Alibaba Cloud D−; xAI, DeepSeek and Mistral failed.
The failing grades spanning the US, China and Europe usefully kills any story about this being a jurisdictional problem. And the methodology choice is the important one — on a curve somebody always looks fine, which is exactly how an industry avoids noticing that its leader is not close to the standard either.
The commitment that came off
No company announced reversing its military-use prohibition. It happened through acceptable-use revisions, widened exception categories and reclassified defence-adjacent work. The endpoint is uniform: every major Western lab now pursues defence partnerships or is open to them.
The restrictions held for exactly as long as they were cheap.
The commercial pressure is not mysterious. Defence budgets are large, terms are long, and sovereign customers do not churn. When the sector's economics look like $5.7bn of quarterly cash burn, a buyer with a decade-long procurement horizon stops being a temptation and becomes a financing strategy.
The finding is procedural
This is not an argument that defence work is wrong. It is an observation about the mechanism: a voluntary commitment is a promise to hold a position until holding it costs something. Every one of these came off under pressure, on roughly the same timeline, without announcement.
Which is the strongest available argument for putting the commitments that matter somewhere they cannot be edited by a policy team.
Future of Life Institute — AI Safety Index — Summer 2026 → · Future of Life Institute — AI Safety Index Summer 2026 (PDF) → · TechTimes — AI Safety Grades Are In: No Lab Tops C+ →