The case for targeted regulation

Source: Anthropic Date: 2024-10-31 URL: https://www.anthropic.com/news/the-case-for-targeted-regulation

Summary

Anthropic published a policy argument for narrowly-targeted AI regulation focused on catastrophic risks, urging governments to act within 18 months. Proposed RSPs as both voluntary industry practice and prototype for enforceable regulation. Three regulatory principles: transparency requirements for RSP-like policies, incentives for safety practices, and surgical simplicity to avoid hampering innovation. Cited Claude models improving from 1.96% to 49% on a cybersecurity benchmark as evidence of rapid capability progress warranting urgency.

Implications

• Safety/policy posture thread. “18 months to act” in October 2024 maps to approximately April 2026 — which is now. This document is a useful benchmark against what actually happened in AI regulation during that period.
• RSPs as regulation template. Anthropic explicitly frames its own RSP as the prototype for enforceable government regulation. If regulators adopt RSP-like requirements, Anthropic’s compliance cost is already paid — competitors bear new costs.
• 1.96% → 49% cybersecurity benchmark. This is one of the most dramatic capability improvement claims Anthropic has made publicly. The specific benchmark (likely a CTF or vulnerability discovery eval) being cited as justification for regulatory urgency is Anthropic saying “look how fast this is moving.”
• “Window for proactive risk prevention is closing fast.” This is the urgency argument that appears in every Anthropic policy submission — calibrated to create pressure for regulatory action before the next capability jump. The tension is that it’s also in Anthropic’s interest to be regulated (RSP as law) before less safety-conscious competitors catch up.
• Watch: what regulation actually passed in the 18-month window; whether the cybersecurity benchmark (1.96% → 49%) was verified by independent researchers; how the “targeted vs. broad regulation” framing evolves in subsequent submissions.