24 February 2026 · Anthropic
Anthropic drops its pledge to stop training if safeguards fall short: Responsible Scaling Policy v3
In February 2026 Anthropic replaced the central promise of its Responsible Scaling Policy: since 2023 it had said it would not train a model unless it could guarantee in advance that its safety measures were adequate. Version 3.0 swaps that hard rule for goals the company declares publicly and grades itself. Time, Bloomberg and others reported it on 24 and 25 February, and Anthropic says the change was needed to keep doing safety work while rivals race ahead.
By Mara Masaeva · maramasaeva.comUpdated 19 November 2026
ArgumentConfirmedMore than one independent source, or a primary document.
What happened
The original policy came out in September 2023. Its core was a commitment never to train an AI system unless safety measures were guaranteed adequate beforehand, which implied pausing if capabilities outran safeguards. Version 3.0, effective 24 February, replaces that with what Anthropic calls non-binding but publicly declared goals, a Frontier Safety Roadmap, and risk reports published every three to six months with external expert review.
PC Gamer notes that the explicit references to pausing disappear in favour of softer language about responsible development, risk management and iterative deployment. Anthropic's own reasoning, quoted by PC Gamer, is that a pause by one company alone would leave the pace to those with the fewest safeguards, and responsible developers would lose the ability to do safety research. Chief science officer Jared Kaplan told Time the change was made in the name of safety rather than competition.
The timing drew attention. Engadget connects it to the same day as the news about Defense Secretary Hegseth and the Pentagon's pressure campaign. A dev.to analysis notes that the two are officially separate, and that Anthropic kept two red lines against the Pentagon: no autonomous weapons and no mass surveillance of Americans.
What came before
Anthropic's own draft bill from October and its listing in November came after this change. The MisAnthropic Statement builds on it.
What I do not know
Sources disagree on the date. Most put the announcement on 24 or 25 February. One outlet, Let's Data Science, dates v3 on 1 April. The February date has more sources behind it.
Whether the change weakened safety or made it more honest is argued, not settled. Critics say the old rule was the point, and Anthropic says it was unworkable when no one else followed it.
The new Risk Reports, and what they contain, are not read for this entry.
Read next
Sources
- NYU Shanghai RITS: Anthropic drops flagship safety pledgepress · main source
Source for the original pledge, the February announcement and the content of the new policy.
- CP24 / CNN: Anthropic ditches its core safety promise in the middle of an AI red line fight with the Pentagonpress
25 February 2026. Source for the nonbinding framework and Kaplan's comment to Time.
- Engadget: Anthropic weakens its safety pledge in the wake of the Pentagon's pressure campaignpress
Source for the timing alongside the Hegseth news.
- PC Gamer: Anthropic ditches its defining safety promisepress
Source for the softer wording in v3 and Anthropic's own reasoning, quoted from the policy document.
- dev.to: The Yieldargument
Analysis post. Source for the effective date of 24 February, the Frontier Safety Roadmap and the two red lines kept against the Pentagon.
- Semafor: Anthropic eases AI safety restrictions to avoid slowing developmentpress
25 February 2026. Independent confirmation of the binary trigger that was removed.
- Let's Data Science: Anthropic revises Responsible Scaling Policy v3press
The one source that dates v3 on 1 April 2026. Kept here because the dates disagree.