Watchtower

Anthropic updated its Responsible Scaling policy from v3.3 to v3.4. The previous wording required observing or expecting double the rate of progress relative to the rate expected in the absence of significant AI contributions to AI R&D. V3.4 now requires doubling relative to both that expected rate and the fastest rate of extended progress previously observed in the absence of significant AI contributions, with a new footnote defining "extended" as spanning at least three model generations. A related clause was changed from trends that "could" lead to greater acceleration to trends that "seem likely to" do so. The stated rationale is that under v3.3, the threshold could trigger even if overall progress were flat or slowing, so long as progress was estimated to be dramatically faster than it would be without advanced AI tools, which Anthropic argues does not indicate the onset of recursive self-improvement, since deceleration from non-AI factors could plausibly continue to offset acceleration from AI-driven gains. The mitigations tied to the threshold are unchanged.

V3.4 also includes multiple changes to Risk Reports. The first change is to its scope, a report will now cover the risks of Anthropic’s models and actions as of a given coverage date, which will be no more than 30 days prior to the reports publication. This includes both publicly deployed models and internally deployed models that meet specific requirements. The second change was around redactions in the Risk Reports, adding a clause that they also aim to minimize redactions of the versions of the report shared with “all regular-clearance Anthropic employees.” Anthropic also added that they would disclose the existence of each redaction made in the public version of the report and aim to give a brief justification for such redactions. v3.4 also adds a clause that Risk Reports will be shared with “at least one external reviewer, but we may choose to split the review process across different external reviewers based on subject matter expertise.” Finally, v3.4 changes who sees the final, unredacted Risk Reports; they now set a number, at least 200 Anthropic employees. They will also share a minimally-redacted version with all of Anthropic’s regular-clearance staff. Previously, all Anthropic’s regular-clearance staff had visibility into the fully unredacted reports. Anthropic discloses and justifies this stating that some information may be both material to the risk assessment and sensitive enough to merit greater internal compartmentalization, especially given the company’s ongoing growth.

See Anthropic's redline v3.4.