Anthropic has revised the internal guidelines it uses to decide how far it can push its AI systems before the risks become too serious to ignore.
The update is the third version of its Responsible Scaling Policy which is a set of self-imposed rules the company created to govern its own AI development.
The update, published February 24, reflects two and a half years of experience with the original policy and a candid admission that some of its early ambitions didn’t pan out.
The biggest change is a clearer split between what Anthropic promises to do itself and what it thinks the broader AI industry should do. The company admitted that one organization was never going to solve catastrophic AI risks on its own.
The original 2023 policy was built on a simple trigger system: if a model became capable enough to pose a serious danger, like helping someone build a biological weapon, tougher safety rules would kick in.
That logic still holds, but Anthropic says it keeps running into a grey area it calls the “zone of ambiguity”. Models would get close to those danger thresholds without clearly crossing them, which made it difficult to justify stronger action.
READ: Cybersecurity Stocks Drop After Anthropic Launches Claude Code Security
Government response also moved more slowly than the company anticipated, and the policy environment has, at times, leaned toward competitiveness over caution.
This update adds two things. First, a public list of safety goals covering areas like security, model behavior, and government policy, with Anthropic committing to report openly on its progress.
Second, a requirement to publish detailed risk assessments every three to six months explaining what each model can do, what dangers that might pose, and what safeguards are in place. In some cases, independent outside reviewers will check these reports.
The policy update is voluntary in that there’s no regulator requiring it, and its influence depends largely on whether other companies follow suit and whether governments eventually build enforceable frameworks around similar principles.
Anthropic’s earlier policy did appear to prompt comparable efforts from OpenAI and Google DeepMind, though how much weight those frameworks carry in practice remains an open question.
As the company put it in the announcement: “This third revision amplifies what worked about the previous RSP, commits us to more transparency about our plans and our risk considerations, and separates out our recommendations for the industry at large from what we can achieve as an individual company.”
