Read an AI provider's safety policy as a deployer's document
Anthropic's revised Responsible Scaling Policy adds regular risk reports and acknowledges that capability thresholds are hard to apply. Buyers still need their own controls.
Anthropic published version 3.0 of its Responsible Scaling Policy on February 24, 2026. It adds risk reports to be published every three to six months, a public Frontier Safety Roadmap, and external review of risk reports in certain circumstances.
The company also wrote that capability thresholds proved harder to apply than expected, describing a zone of ambiguity in which it is unclear whether a model has crossed a threshold.
ESS view: provider safety policies are useful disclosures. For an agency deploying AI, the controls that protect operations are its own: access permissions, logging, rollback and evaluation on its actual tasks.