⚡ Quick Summary
Published by Anthropic, this policy sets out the company's voluntary framework for managing catastrophic risks from advanced AI systems. It is the third iteration of the Responsible Scaling Policy (RSP), marked version 3.0 and effective February 24, 2026, and it states how Anthropic identifies and evaluates risks, decides on development and deployment, and aims to ensure the benefits of its models exceed their costs.
The central change in this version is the separation of the company's own plans from more ambitious industry-wide recommendations, driven by what the document calls a collective action problem: ecosystem risk depends on the actions of many developers, so Anthropic will not commit unilaterally to the industry-wide column. Section 1 presents a three-column table mapping capability or usage thresholds — non-novel chemical/biological weapons production, novel chemical/biological weapons production, high-stakes sabotage opportunities, and automated R&D in key domains — to Anthropic's planned mitigations and to its recommendations for frontier developers, including security roughly in line with RAND SL4 at higher thresholds.
New requirements include Frontier Safety Roadmaps setting goals across Security, Alignment, Safeguards and Policy, and Risk Reports published every 3–6 months and subject to external review, alongside governance commitments, competitor-contingent commitments and a public changelog.
🧩 What’s Covered
The document is organised into an introduction, four numbered sections, two appendices and a changelog.
- Introduction and scope: defines the RSP as a voluntary framework for catastrophic risks, explains the shift away from committing to absolute risk reduction, and notes that the RSP is one part of Anthropic's safety approach, with regulatory requirements handled in separate documents.
- Recommendations for industry-wide safety (Section 1): a three-column table of four thresholds — non-novel CBRN weapons production, novel CBRN weapons production, high-stakes sabotage opportunities, and automated R&D in key domains — each paired with Anthropic's planned mitigations (for example maintaining ASL-3 protections) and industry-wide recommendations (for example RAND SL4-level security).
- Frontier Safety Roadmap (Section 2): ambitious but achievable goals for improving risk mitigations, shared with employees, the Board and the Long-Term Benefit Trust, published in redacted form, with past roadmaps kept available.
- Risk Reports (Section 3): scope and timing (all publicly deployed models and qualifying internal models; publication every 3–6 months), general expectations, required factual contents and risk analyses, procedures ending in CEO and Responsible Scaling Officer approval, and publication with redactions.
- External review (Section 3.6): reviewer selection criteria, timing and access, the contents of the review, and assessment of redaction scope, justification, balance and materiality, with public commentary by reviewers.
- Governance (Section 4): seven measures covering the RSO role, internal transparency, noncompliance reporting, employee agreements, internal review, an annual third-party procedural compliance review, and approval of policy changes by the Board in consultation with the LTBT.
- Appendices A and B and the changelog: competitor-contingent commitments for three scenarios, notes on why AI Safety Levels no longer list specific controls for future capability levels, and version history from RSP v1.0 (September 19, 2023) to v3.0.
💡 Why it matters?
The policy shows how a frontier developer turns risk thresholds into operational commitments: capability thresholds trigger documented mitigations, Risk Reports must state absolute and marginal risk, and external reviewers assess both the reasoning and the redactions. That makes it a reference point for auditors and board members judging whether safety claims are evidenced, and for policy analysts comparing voluntary commitments with statutory regimes. The document itself notes that the RSP may serve some regulatory requirements but is not comprehensive, and that where requirements exceed or differ from it, they are addressed through separate compliance frameworks, naming California SB 53.
❓ What’s Missing
The document states that it cannot presently give highly specific advance detail on which evaluations determine whether risk thresholds have been passed or what mitigations will be needed, so thresholds are expressed as arguments for safety rather than fixed tests. The current Frontier Safety Roadmap and the ASL-3 controls are referenced but not reproduced. Risk Reports will be published in redacted form for legal compliance, intellectual property, public safety and privacy reasons. External review is described as an experiment, as no established organisations or procedures exist, and regulatory compliance mapping is deliberately excluded.
👥 Best For
Frontier-lab safety and governance teams drafting risk-threshold policies; policy analysts, auditors and board or trust members assessing voluntary AI safety commitments; and researchers examining how a developer converts capability thresholds into mitigations, external review and accountability mechanisms.
📄 Source Details
Responsible Scaling Policy, version 3.0, published by Anthropic; no individual authors are named. Effective February 24, 2026; 19 pages; English. The changelog lists earlier versions from RSP v1.0 (September 19, 2023) up to the current v3.0. The document prints the address www.anthropic.com/responsible-scaling-policy and a roadmap link at anthropic.com/responsible-scaling-policy/roadmap. The text extraction covered all 19 pages.
