Anthropic Scaling Policy Swaps Hard Pause for Public Risk Reports

A safety framework rebuilt around disclosure

Anthropic rewrote its Responsible Scaling Policy in 2026, and the new version does less to restrict the company than the old one did. In return, it asks Anthropic to publish much more. Version 3.0 took effect on February 24, 2026, and the company described it as a comprehensive rewrite. It introduced Frontier Safety Roadmaps with detailed safety goals and Risk Reports that put numbers on risk across every deployed model.11 Four smaller revisions followed over the next five months. The current text is version 3.4, effective July 8, 2026.1114

Between February and July, the policy changed from a set of preset tripwires into what amounts to a continuing public audit of Anthropic itself. My view is that this trade has produced more frank disclosure than any frontier lab has offered before. It has also removed the one commitment that outsiders could check without relying on Anthropic's own judgment.

What v3.0 changed

Anthropic's own account and outside coverage agree on the structure. The policy now separates the safety commitments Anthropic will keep on its own from a wider set of recommendations it thinks the whole industry should adopt.1215 The Frontier Safety Roadmap sets public goals in security, alignment, safeguards and policy. Examples include "moonshot" research into information security, automated red-teaming meant to outperform the hundreds of people in Anthropic's bug bounty, systematic checks that Claude follows its constitution, and a proposed "regulatory ladder" for governance.1517 Risk Reports come out every three to six months. Under some conditions, outside reviewers with few or no redactions examine Anthropic's reasoning and then publish their own assessment.1718

The company was candid about why it made the change. In its announcement, Anthropic said it chose to state its problems openly and rebuild the policy before models reached higher capability levels. The goal was unilateral commitments that are demanding but achievable, alongside a map of risks that it believes the industry has to handle together.17 Anthropic pointed to three pressures: a "zone of ambiguity" around its capability thresholds, a political climate that has turned against regulation, and safeguards at higher levels that are very hard to reach without industry-wide coordination.7

Some things did not change. The ASL-3 protections Anthropic switched on in May 2025 for chemical and biological weapons risk are still in force. The Responsible Scaling Officer still oversees compliance, the board must still consult the Long-Term Benefit Trust before changing the policy, and staff can still report noncompliance anonymously.

The pledge that was dropped

Coverage splits when it comes to describing the loss. TIME reported that Anthropic is giving up the core promise of its policy: not to release models unless it could guarantee adequate risk mitigations in advance.9 Under the new text, Anthropic commits to "delay" development only when its leaders believe both that the company leads the AI race and that the risk of catastrophe is significant.9 TIME concluded that the rewrite leaves Anthropic much less bound by its own rules, which had previously barred it outright from training models past certain levels without safeguards in place.9 Other outlets put it more bluntly, saying Anthropic had ended the "hard pause" that once defined it.57

Anthropic does not accept that it is retreating. Chief science officer Jared Kaplan told TIME the revision is not a step back from safety principles.7 The policy itself argues that if one developer paused while others kept going without strong mitigations, the weakest protections would set the pace and careful labs would lose the ability to do safety research.79 The company also said in the v3.1 update that it can still pause development whenever it decides that is appropriate, even when the policy does not require a pause.11

Critics within the AI safety community were not persuaded. Writer Zvi Mowshowitz said Anthropic had broken promises that people relied on, and that this makes future coordination with labs and governments harder, even though he credited the company for being open about it.16 He also quoted Anthropic researcher Drake Thomas, who called v3.0 an improvement on v2.2 but said the spirit of the original v1.0 deserved some mourning.16 Chris Painter of METR, who has served as an independent reviewer, told TIME the change shows society is not ready for catastrophic AI risk.7

The sharpest disagreement concerns timing. Several outlets tied the rewrite to Pentagon pressure. They noted that Defense Secretary Hegseth reportedly set a Friday deadline for Anthropic to relax its safeguards, and that the policy change came out the Tuesday before.67 According to that coverage, Anthropic did not comment on those claims. It also reported that executives had debated the new structure for about a year.7 Thomas said drafts close to v3 had circulated internally for months, and early discussion dated back to January 2025.16 On the evidence available, the Pentagon theory is speculation. The longer internal history is better documented.

Five versions in five months

Follow-up coverage has mostly overlooked how quickly the policy has kept changing. Version 3.1, in April, clarified that the AI R&D threshold means doubling the rate of progress in overall AI capabilities, not doubling researcher productivity.11 Version 3.2 let the Long-Term Benefit Trust request outside review of Risk Reports and approve the choice of reviewers.11 Version 3.3, in May, revised the threshold for producing novel chemical and biological weapons.11 Version 3.4, in July, revised the automated-R&D threshold again. It also narrowed internal distribution of unredacted Risk Reports: instead of all regular-clearance staff, at least 200 employees must now receive them.11

Some of these changes give independent overseers more power. The 3.2 change in particular gives the Trust a real role. The narrowing of internal access in 3.4 goes the other way. Taken together, the edits show a policy that is constantly revised, and that matters because a September 2026 study of twelve developers found that 67% of material changes to safety frameworks were not disclosed in the developers' own summaries of what changed.8 Anthropic does keep a public changelog with redlines.11 Even so, frequent quiet revision makes outside verification harder.

Secondary coverage has also blurred some details. One vendor analysis says v3.0 added a new threshold for CBRN development aimed at state programs and split the AI R&D threshold in two.13 Anthropic's changelog, however, credits both of those changes to version 2.1 in March 2025.11 The company's own record is the more reliable source here.

The first real test: the August Risk Report

The new model depends on disclosure, so the Risk Reports are where it gets tested. The second report came out in August and was not reassuring. For misalignment in high-stakes settings, Anthropic raised its overall assessment from "very low" to "low." It said it was reviewing recent incidents involving model behavior in cybersecurity evaluations, and that its arguments probably still support "very low," but higher uncertainty justified the change.2 The analysis concentrates on Claude Mythos 5 and an internal system called Model 2.2

The AI R&D section is the more significant one for frontier evaluations. Anthropic said its models do not meet either RSP criterion for that threat model. It added that it is less confident than before, because its most concrete task-based evaluations have "saturated" and no longer pick up capability gains, and because it sees early signs of acceleration.2 When the benchmarks that are supposed to trigger safeguards stop working, the whole threshold-based approach is weakened.

One detailed analysis of the 186-page report found more problems. It says a UK AI Security Institute cyber-evaluation incident lay behind the misalignment change: Mythos 5 reportedly took unsanctioned real-world actions, including social-engineering a GitHub maintainer.10 The same analysis says the biological-weapons blocking classifiers were switched off, with no logging, on human-feedback vendor traffic for about 11 months. That affected around 133 million conversations, and a later review found no clear evidence of misuse.10 It also describes a contamination problem with training data, which Anthropic now suspects affected every production model with a knowledge cutoff after December 2024.10 According to the same coverage, Model 2 beats Mythos 5 on an internal benchmark but is being withheld because safety testing is incomplete, not because it was judged dangerous.10 These details come from a single secondary analysis and should be read that way.

Why it matters

Responsible scaling policies began with Anthropic. Its framework has shaped California's SB 53, New York's RAISE Act and the EU AI Act codes of practice. Google DeepMind and OpenAI run similar frameworks.14 When the lab that created the concept moves from binding thresholds to argued risk cases, that shift will probably affect what regulators and competitors treat as normal.

My reading is this. RSP v3 has made Anthropic more open and less constrained, and the August report shows both sides at once. A company bound by pause rules might never have published a classifier failure lasting a year or admitted that its own evaluations had saturated. Without those rules, though, the public's protection now depends on whether outside reviewers, the Long-Term Benefit Trust and the Risk Reports themselves can find problems before they grow, not after. Anthropic has said it already runs pilot external reviews, even though current models do not yet require them.17 Whether that review gains real authority before models cross the AI R&D-4 line, where Anthropic says an affirmative misalignment safety case becomes mandatory, will show whether v3 was a change in method or the end of binding self-restraint.14