Anthropic Unveils Third Version of Responsible Scaling Policy

Anthropic, a leading AI company, has recently launched the third iteration of its Responsible Scaling Policy. This release marks a significant milestone in the company’s approach to mitigating catastrophic risks associated with artificial intelligence. The new policy comes after an extensive period of real-world testing spanning two and a half years.

The latest update, announced on February 24, 2026, introduces three key changes to the policy framework. These changes include a clear distinction between Anthropic’s individual commitments and industry-wide recommendations, the introduction of a new Frontier Safety Roadmap with public accountability metrics, and the implementation of mandatory external review of Risk Reports under specific conditions.

Key Updates and Enhancements

One of the most notable changes in the new Responsible Scaling Policy is Anthropic’s acknowledgment that certain safety measures cannot be achieved by a single company alone. The policy now explicitly delineates between safeguards that Anthropic can implement independently and those that require collaborative industry efforts.

A recent RAND report referenced by Anthropic highlights the challenges in achieving “SL5” security standards, indicating that external assistance, potentially from the national security community, may be necessary.

Instead of diluting the requirements to facilitate compliance, Anthropic opted for a complete restructuring of the policy. The updated framework now outlines two distinct tracks: commitments that the company will fulfill irrespective of external factors and recommendations that Anthropic believes the entire AI industry should adopt.

Evaluating Past Performance

Anthropic’s retrospective analysis of the previous versions of the Responsible Scaling Policy offers valuable insights. The policy successfully instilled a safety-first approach within internal teams, prompting competitors like OpenAI and Google DeepMind to adopt similar frameworks swiftly. Notably, ASL-3 safeguards were effectively implemented in May 2025.

However, challenges arose with the ambiguity of capability thresholds. For instance, the assessment of biological risks posed difficulties as models passed initial tests but lacked definitive results. Additionally, shifting policy priorities towards AI competitiveness and economic growth hindered federal safety discussions.

Introducing New Accountability Measures

The Frontier Safety Roadmap introduces specific, publicly-graded objectives, including ambitious R&D projects for information security, advanced automated red-teaming systems, and comprehensive documentation of critical AI development activities. These records will be analyzed by AI for potential insider threats.

Risk Reports will now be published every 3-6 months, detailing the alignment of capabilities, threat models, and mitigations. External reviewers with access to unredacted or minimally-redacted information will publicly critique Anthropic’s rationale. Despite current models not triggering the external review requirement, the company has initiated pilot programs.

Implications for the AI Industry

The restructuring of the Responsible Scaling Policy coincides with increased scrutiny on AI governance frameworks. Legislative measures such as California’s SB 53, New York’s RAISE Act, and the EU AI Act’s Codes of Practice now mandate frontier developers to disclose catastrophic risk frameworks.

Anthropic’s approach of segregating unilateral commitments from industry recommendations may set a precedent for the sector. By acknowledging the limitations of voluntary self-regulation and advocating for coordinated government action, Anthropic positions itself as a leader in responsible AI development.

Ultimately, Anthropic’s transparent acknowledgment of the collaborative nature of addressing AI risks may have a profound impact on the industry, emphasizing the importance of collective efforts in ensuring AI safety.

Image source: Shutterstock