Cambridge researchers propose concrete AI safety framework with government response plans

Cambridge Safety Framework Proposes Concrete Government Preparedness for AI Incidents

Cambridge researchers have moved beyond abstract AI safety talk to propose measurable frameworks and government contingency planning for high-impact AI scenarios.

The core tension in existing safety proposals is vagueness masquerading as substance. Labs like Anthropic and OpenAI mention that AI is accelerating their own R&D cycles, but they report this using incompatible metrics—OpenAI frames it as "intern level" capability, Anthropic as "Claude spinning up a thousand ideas"—making it impossible to compare actual progress across the industry. The Cambridge proposal cuts through this by requiring standardized monitoring of what percentage of AI R&D is conducted by AI systems themselves, with firm metrics on both code and dollar allocation. The financial incentive to stay vague is real: being further along the recursive self-improvement curve carries competitive advantages, so neither lab has strong external accountability on where they actually stand relative to each other.

The second lever is government preparedness. The researchers argue that policy should move past "we need to talk about this" and into specific response planning. They propose that the government develop contingency playbooks for discrete, plausible failure modes: a rogue agent conducting a cyberattack and taking down critical internet infrastructure, biosecurity incidents, economic dislocation from sudden labor displacement. The framing borrows from existing crisis response architecture—FEMA has concrete plans for hurricanes, evacuation routes, resource distribution. The proposal does not prescribe solutions yet, but insists the government should have tabletop scenarios and response frameworks ready, similar to how disaster simulation is practiced in other domains.

The proposal includes an executive summary and full paper described as accessible to non-specialists, signaling an attempt to move the conversation from academic margins into policy channels.