Announcing our updated Responsible Scaling Policy

Anthropic has overhauled its Responsible Scaling Policy, the internal rulebook that decides when a Claude model is too dangerous to ship without extra safeguards. The new version, published today, keeps the company's basic promise intact — no training or deploying models without adequate protections — but swaps out some of the original framework's rigidity for something closer to a living document.

The headline change is a pair of specific capability thresholds. If a model can meaningfully help someone with basic technical skills build a chemical, biological, radiological or nuclear weapon, that triggers ASL-3 safeguards: tighter access controls, better protection of model weights, and layered deployment monitoring including red-teaming before launch. If a model can autonomously handle complex AI research tasks that normally require human PhDs, that's treated as an even bigger deal, potentially requiring ASL-4 protections, because a model that can improve AI itself could accelerate progress faster than anyone can track.

Anthropic also came clean about stumbling in year one. A self-review found the company missed one evaluation deadline by three days, was fuzzy about documenting changes to placeholder tests, and in a few cases probably could have squeezed better performance out of models using standard tricks like chain-of-thought prompting before declaring them safe. None of this, the company says, put anyone at meaningful risk — the delayed evaluations turned out more accurate than the ones they replaced, and models stayed comfortably below the danger thresholds throughout. But it's a rare admission that even the AI safety and security policy is a work in progress.

The reshuffle extends to personnel too. Jared Kaplan, Anthropic's co-founder and chief science officer, is taking over as the company's Responsible Scaling Officer from CTO Sam McCandlish, who ran point during the policy's first year. Anthropic is also hiring a Head of Responsible Scaling to coordinate the growing list of teams — Frontier Red Team, Trust & Safety, Alignment Science, and others — now tied into RSP compliance, plus soliciting outside feedback from AI safety institutes and independent experts on how it grades its own models.