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 major 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 adjustments: 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 assessment of Threat Studies underneath sure situations.
What Truly Modified
Essentially the most notable shift? Anthropic is now brazenly 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 not possible.
A RAND report cited by Anthropic states that “SL5” safety requirements aimed toward stopping top-tier cyber threats are “at the moment not potential” and “will probably require help from the nationwide safety group.”
Moderately than water down these necessities to make compliance simple, 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 the complete AI {industry} must undertake.
The Sincere Evaluation
Anthropic’s autopsy on RSP variations 1 and a pair of is refreshingly candid. What labored: the coverage compelled inside groups to deal with security as a launch requirement, and opponents like OpenAI and Google DeepMind adopted related frameworks inside months. ASL-3 safeguards have been efficiently activated in Could 2025.
What did not work: functionality thresholds proved much more ambiguous than anticipated. Organic danger evaluation gives a telling instance—fashions now move most fast checks, making it arduous 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 atmosphere 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” initiatives for data safety, automated red-teaming techniques that exceed present bug bounty contributions, and complete data of all crucial AI growth actions—analyzed by AI for insider threats.
Threat Studies will publish each 3-6 months, explaining how capabilities, menace 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 assessment requirement.
Business 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 Follow have all begun requiring frontier builders to publish catastrophic danger frameworks—necessities Anthropic addresses via its current Frontier Compliance Framework.
Whether or not opponents comply with Anthropic’s lead on separating unilateral commitments from {industry} suggestions stays to be seen. The strategy basically acknowledges that voluntary self-regulation has limits, whereas positioning the corporate to advocate for coordinated authorities motion with out showing to demand guidelines it will probably’t comply with itself.
For the broader AI sector, Anthropic’s clear acknowledgment of what single corporations can’t obtain alone could show extra influential than the technical coverage particulars themselves.
Picture supply: Shutterstock
