Evaluating AI risk only by today’s public open models is no longer enough. The lag matters: how quickly capabilities from closed frontier systems move into cheaper, available, reproducible models.
The UK AI Security Institute says autonomous AI cyber capability is advancing quickly and the gap between closed frontier models and open-weight systems on cyber capability benchmarks is shrinking. FT highlighted the same line: capabilities recently limited to the closed top of the market start reaching open models after only a few months.
What changed
In the old picture, open-weight models could be treated as a delayed lower-risk layer: frontier moves first, open models catch up much later, and organizations have time to adapt policy, detection, and safeguards.
Now the lag is shorter. That does not mean every open-weight model is automatically dangerous. It does mean capability forecasting must track not only the absolute level today, but the speed of capability transfer.
Why it matters for products
Cyber capability is different from an ordinary benchmark race. If a model writes better code or reasons better, it improves products. If a model autonomously finds and exploits vulnerabilities, it can improve defensive tooling while also lowering the cost of attack workflows.
Product risk models should track:
- which capabilities frontier systems already have;
- how long it takes similar capabilities to appear in open-weight models;
- which tools make an agentic cyber workflow possible around the model;
- which actions the model can perform without an outside specialist;
- which safeguards live in the model and which live only in a closed API layer.
New Runtime Read
A risk model must account not only for current open models, but for the short lag that moves frontier capabilities into cheap and accessible systems.
If you build an AI product with access to code, infrastructure, pentest data, or internal secrets, update the threat model by capability shifts, not once a year. The open-weight layer is becoming a mass channel for capability diffusion, not an experimental edge.