Capability that helps and harms with equal facility. The same chemistry knowledge that supports research supports misuse; the same code ability that finds vulnerabilities to fix finds them to exploit. It means capability restriction is a blunt instrument, and much of the difficulty in AI safety policy comes from this.
Technical view
Concentrated in cyber, chemical, biological, radiological and nuclear domains, and increasingly in autonomous capability. Policy responses include capability evaluation before release, tiered access with know-your-customer requirements, refusal training for specific request classes, and staged release. The counter-argument is that restriction imposes real costs on legitimate research while determined actors have alternatives, which is why marginal uplift rather than absolute capability is the measure used.
Related
Part of AI Terminology Index, section: Security and Privacy.