Tony Kim
Feb 24, 2026 20:48
Anthropic releases third model of Accountable Scaling Coverage, separating firm commitments from industry-wide suggestions after 2.5 years of testing.
Anthropic has launched the third iteration of its Accountable Scaling Coverage, marking a big restructuring of how the AI firm approaches catastrophic danger mitigation after two and a half years of real-world implementation.
The replace, revealed February 24, 2026, introduces three main modifications: a transparent separation between what Anthropic can obtain alone versus what requires industry-wide motion, a brand new Frontier Security Roadmap with public accountability metrics, and necessary exterior evaluate of Danger Experiences underneath sure circumstances.
What Really Modified
Essentially the most notable shift? Anthropic is now overtly admitting that some security measures merely can’t be applied by a single firm. The earlier RSP’s higher-tier safeguards (ASL-4 and past) have been left deliberately imprecise—seems that wasn’t simply warning, it was as a result of reaching them unilaterally could also be inconceivable.
A RAND report cited by Anthropic states that “SL5” safety requirements geared toward stopping top-tier cyber threats are “at the moment not potential” and “will probably require help from the nationwide safety group.”
Slightly than water down these necessities to make compliance straightforward, Anthropic selected to restructure fully. The brand new RSP now explicitly maps out two tracks: commitments the corporate will meet no matter exterior elements, and suggestions it believes your complete AI {industry} must undertake.
The Trustworthy Evaluation
Anthropic’s autopsy on RSP variations 1 and a couple of is refreshingly candid. What labored: the coverage compelled inner groups to deal with security as a launch requirement, and rivals like OpenAI and Google DeepMind adopted comparable frameworks inside months. ASL-3 safeguards have been efficiently activated in Might 2025.
What did not work: functionality thresholds proved way more ambiguous than anticipated. Organic danger evaluation gives a telling instance—fashions now cross most fast checks, making it laborious to argue dangers are low, however outcomes aren’t definitive sufficient to show dangers are excessive both. By the point wet-lab trials full, extra highly effective fashions have already shipped.
The political setting hasn’t helped. Federal safety-oriented discussions have stalled as coverage focus shifted towards AI competitiveness and financial progress.
New Accountability Mechanisms
The Frontier Security Roadmap introduces particular, publicly-graded targets together with “moonshot R&D” tasks for info safety, automated red-teaming programs that exceed present bug bounty contributions, and complete information of all vital AI growth actions—analyzed by AI for insider threats.
Danger Experiences will publish each 3-6 months, explaining how capabilities, risk fashions, and mitigations match collectively. Exterior reviewers with “unredacted or minimally-redacted entry” will publicly critique Anthropic’s reasoning.
The corporate is already working pilots regardless of present fashions not but triggering the exterior evaluate requirement.
Trade Implications
This restructuring arrives as AI governance frameworks face growing scrutiny. California’s SB 53, New York’s RAISE Act, and the EU AI Act’s Codes of Apply have all begun requiring frontier builders to publish catastrophic danger frameworks—necessities Anthropic addresses by way of its current Frontier Compliance Framework.
Whether or not rivals comply with Anthropic’s lead on separating unilateral commitments from {industry} suggestions stays to be seen. The method primarily acknowledges that voluntary self-regulation has limits, whereas positioning the corporate to advocate for coordinated authorities motion with out showing to demand guidelines it might probably’t comply with itself.
For the broader AI sector, Anthropic’s clear acknowledgment of what single corporations can not obtain alone could show extra influential than the technical coverage particulars themselves.
Picture supply: Shutterstock
