The Cyber Risk Discourse is Broken

The latest fight over open AI models, according to Interconnects, is badly skewed. On one side are people treating open weights as an unacceptable cyber threat. On the other are people who say banning them would leave the world less safe. The author’s complaint is that both camps are skipping the messy part: what actually happens to cyber risk if open models are restricted while closed models keep advancing.

That matters because the public record, the essay says, points in a different direction than the panic. Closed models have already been documented as the cause of most existing cyber attacks. The piece leans on that to make a blunt point: the usual “open dangerous, closed safe” story may be too neat. A more accurate version could be “open unsafe, closed unsafe,” with the difference coming down to who can access the tools, how strong the models are, and how porous the safeguards turn out to be.

The author also pushes back on a recent Anthropic report about GLM-5.3 as an offensive cyber tool. The technical work is described as broadly reasonable. The problem, in the author’s view, is that it doesn’t answer the bigger policy question of what a ban would do in practice, or why Chinese companies think their own releases are safe enough. That, the essay argues, is the real puzzle. Every policy move changes the shape of the risk, but almost never in a clean way.

That logic leads to a fairly sharp conclusion: if someone wants to ban the newest open-weight models to slow cyber diffusion, then public-facing APIs for frontier closed models should probably also be illegal. Otherwise, the offense-defense gap could widen, because closed models may improve faster than guardrails do. The author thinks open weights may still be the better near-term tool for defenders in some settings, including air-gapped government networks where APIs can’t be used.

The China section is equally pointed. The essay says Chinese companies do care about AI safety, but they do so inside a system shaped by political stability, social risk, and government review. Major model releases reportedly have to be registered with the Chinese government, though it’s unclear how far that extends beyond older information-control concerns. The piece also argues that running thorough safety evaluations can cost tens of millions of dollars in compute, which creates a strong incentive to keep the bar lower than critics want. The author’s broad claim is simple: the current debate is too often a loud Western argument about open weights, while the real cyber-risk picture is wider, uglier, and nowhere near settled.