AI Safety Scores Are In: What the 2026 Index Actually Tells Us
The 2026 AI Safety Index and Bengio's report land together. Here's what the scores reveal, where labs fall short, and what teams should do now.
Two documents dropped this summer that most AI teams should read together, not separately. The International AI Safety Report 2026—Yoshua Bengio's hundred-plus-author synthesis, backed by over 30 countries—gives us the scientific picture. The July 2026 AI Safety Index gives us the scoreboard: how leading labs actually perform on safety and security domains. Read side by side, they tell a sharper story than either does alone.
The gap that matters
The research consensus is getting clearer on capability trends and the risk surface that comes with them. The Index, meanwhile, shows that lab performance on those same domains is uneven. That mismatch is the headline. It's not that labs are ignoring safety—most publish frameworks and red-team results now. It's that the evidence behind those claims varies wildly, and the Index is one of the first attempts to grade it consistently across companies.
A rough sketch of what the domains look like:
| Domain | What the Index probes | |---|---| | Risk assessment | Pre-deployment evals, third-party testing | | Security | Model weight protection, insider threats | | Transparency | Disclosure of capabilities and incidents | | Governance | Board oversight, escalation paths |
Why the scores are noisy
Be skeptical of a single number. Safety scoring is young, and the methodologies behind it are still being argued over. A lab that scores well on transparency may be weak on security; one with strong evals may have thin governance. Treat the Index as a diagnostic lens, not a leaderboard. The useful signal is the pattern—where the whole industry clusters low, that's a structural problem, not a company problem.
What this means for your team
If you're building on top of frontier models, the Index is a procurement input, not a press release. Three practical moves:
- Map your dependencies. Know which lab's models sit in your critical paths, then check their Index domains against your own risk register.
- Demand evidence, not frameworks. Ask vendors for third-party eval results and incident disclosure history, not just a policy PDF.
- Build your own evals. The Index grades labs; it doesn't grade your deployment. A small internal red-team suite beats borrowed confidence.
The takeaway
The 2026 report and Index together mark a shift: AI safety is moving from vibes to measurement. That's healthy, even if the measurements are imperfect. The teams that treat these scores as a starting point for their own diligence—rather than a verdict to cite—will be the ones who catch problems before they ship. Start with one domain this quarter. Pick the one you understand least.