Defender’s Dilemma Releasing Dual-Use AI Models as Capabilities Climb

The panel opens by asking whether dual-use AI resembles earlier fights over encryption and export controls. Speakers argue that the central question is who decides which users get access, and they compare AI to telephones, cars and airplanes to show how different mental models lead to different levels of concern. A recurring theme is that friction, including limits on availability, built-in safeguards, legal consequences and defender action, has historically managed dual-use risk.

The middle section covers diffusion. Participants expect capable open-weight models to trail frontier systems by only months, helped by falling token costs and local hardware. They discuss whether coordinated disclosure still works when findings spread instantly, and they point to AI-assisted cryptanalysis and firmware diffing as examples of capabilities that already exist. Alignment, meaning whether a model recognizes when it is in the wrong situation, is raised as a gap in how capability is currently measured.

The closing segment looks at defense. Suggestions include intent-based access decisions, graduated verification instead of yes-or-no checks, audit frameworks that regulators can converge on, and treating integrity as a neglected security property. The practical advice is to press vendors to adopt AI-assisted defense now, prepare for a difficult stretch of roughly eighteen to twenty-four months, and keep contributing ideas to the community.