This is a call to everyone working on AI safety frameworks: testing AI models only one at a time is dangerous.

The companies building the most capable AI models publish safety frameworks. These documents say when a model is too dangerous to release without extra protection. All of them ask the same question: is this model dangerous? I think it is the wrong question, and here is why.

Safe alone, unsafe together

Jones, Dragan and Steinhardt showed at ICML 2025 that “adversaries can misuse combinations of models even when each individual model is safe”. The trick is simple. Split a harmful task into pieces. Give the hard but harmless pieces to a strong model, which is happy to help. Give the easy but harmful pieces to a weak model that never refuses. In one of their experiments, a strong model and a weak open one produced vulnerable code 43% of the time. Neither model alone went above 3%. Nobody had to jailbreak anything.

So each model passes its test, and the pair fails.

What the frameworks do

I have read Google’s Frontier Safety Framework closely, including its latest revision of April 2026 (version 3.1). It defines capability thresholds, and it says that a model below every threshold poses an acceptable risk, because it lacks the capabilities to contribute substantially to severe harm (Section 1.3.5). The model is tested with its own scaffolding and, for misuse, compared with a world without generative AI. In other words, the model is tested as if it were the only one out there.

Other models do appear in the framework, but only after a threshold is reached, and only in examples where they lower the risk: if a similar model is already public, releasing a new model adds little. The opposite case, where another model makes the new model more dangerous, is not there. And the framework names what reopens the verdict on a released model: a new version trained by Google. A new model released by somebody else is not on that list.

Google is not alone. As far as I can tell, the frameworks of OpenAI and Anthropic also assess one model at a time, and none of the three mentions an attacker who combines models.

Curiously, the European Code of Practice already lists “interactions with other AI models and/or AI systems” among the possible sources of systemic risk. The rule book has seen the problem. The frameworks have not.

Why well-meaning companies do this?

I do not think these companies are careless. They are in a race. Caution is costly for whoever shows it first, and no company can make AI safe on its own. The frameworks say so explicitly themselves. Google writes that its security recommendations only work if the whole field applies them. OpenAI keeps the option of lowering its requirements if a competitor releases a risky system without comparable safeguards. Anthropic rewrote its policy in 2026, explaining that safeguards at higher levels “might prove outright impossible to implement without collective action”.

When everybody measures only what their own model adds to what is already out there, everybody’s share looks small. The total does not.

A proposal: share the risk

Mathematics has had a tool for this since 1953: the Shapley value.

Take the set NN of models and tools that people can access today, say there are n=|N|n = |N| of them. For every subset SS of NN, let v(S)v(S) be the risk when exactly the models in SS are available, measured against a world with none of them, so that v(∅)=0v(\emptyset) = 0. You can think of v(S)v(S) as the test that frameworks already run, applied to a group of models instead of one. The share of a model mm is “its average contribution over all the orders in which the models could have arrived”:

φm=∑S⊆N\{m}|S|!(n−|S|−1)!n!(v(S∪{m})−v(S)). \varphi_m = \sum_{S \subseteq N \setminus \{m\}} \frac{|S|! \, (n - |S| - 1)!}{n!} \left( v(S \cup \{m\}) - v(S) \right) .

With two models AA and BB, this is

φA=12v({A})+12(v({A,B})−v({B})). \varphi_A = \frac{1}{2} \, v(\{A\}) + \frac{1}{2} \left( v(\{A, B\}) - v(\{B\}) \right) .

Recall Jones, Dragan and Steinhardt at ICML 2025. Consider v(S)v(S) a success rate instead of risk. Then, according to their experiments, wach model alone succeeds v({A})=v({B})=v(\{A\}) = v(\{B\}) = 3%, and together they reach v({A,B})=v(\{A, B\}) = 43%. Using the Shapley value, each model gets a share between 0.20 and 0.23! Two models that looked harmless now carry roughly half of the problem each.

Three things make this attractive.

  1. The shares add up to the total risk, so no risk is left without an owner.
  2. Models that are interchangeable get equal shares, so a second model that copies the first does not get a free pass.
  3. The order of arrival does not matter, so being the last one to release is not an excuse.

This is not an exotic idea. Researchers at the Bank for International Settlements proposed the Shapley value to attribute systemic risk to individual banks, for a similar reason: the risk of a system is not the sum of the risks of its parts. Anthropic’s policy already speaks of its marginal contribution to the risk of the whole ecosystem. The Shapley value is the version of that idea that adds up.

There are limits, of course. We need to agree on the set NN, and a short list of the strongest public models is a start. Measuring vv on groups of models costs more than measuring it on one, although five models give only 32 groups. And a share is not a verdict: we still need to say how much total risk is acceptable.

What I ask

  • To AI companies: test your models in combination with other models, including your own, and repeat the test when the ecosystem changes.
  • To regulators and standards bodies: the Code of Practice already names the problem, so ask for the test.
  • To researchers: measure how much models complement each other on the tasks that matter, and find cheap ways to estimate these shares.
  • To everybody else: when you read that a model is “below the threshold”, ask “below the threshold together with what?”.