Anthropic RSP 3.2: LTBT Can Force External Risk-Report Reviews
Anthropic Responsible Scaling Policy v3.2 (April 29) lets the Long-Term Benefit Trust request external reviews of Risk Reports, approve reviewer selection, and require regular briefings. The trade is harder self-oversight optics — still voluntary, still company-controlled policy text.
Version bumps on voluntary safety policies rarely change the product. This one changes who can demand a second set of eyes on Anthropic’s own Risk Reports — and who picks those eyes.
Anthropic released Responsible Scaling Policy Version 3.2, effective April 29, 2026. The RSP is the company’s voluntary framework for catastrophic risks from advanced systems. The 3.2 delta concentrates power in the Long-Term Benefit Trust (LTBT) rather than rewriting ASL thresholds.
What changed
| Provision (Anthropic primary) | v3.2 |
|---|---|
| External Risk Report review | LTBT can request external review |
| Reviewer selection | LTBT can approve Anthropic’s choice of external reviewers |
| Briefings | Formal requirement for regular LTBT briefings on RSP implementation |
Full RSP text and a redline vs prior versions are on Anthropic’s site. Preceding cadence: v3.1 effective April 2; v3.0 effective February 24 — those introduced Frontier Safety Roadmaps and Risk Reports.
Claims vs checks
Policy text and redline are Anthropic primary. “Strengthened independent oversight” is the company’s framing; the LTBT remains Anthropic’s designated trust structure, not a government regulator. No third-party audit of whether prior Risk Reports would have changed under 3.2 powers is published with the release.
Limits
- Voluntary framework — no statutory enforcement.
- External review is requestable, not automatically mandatory for every Risk Report.
- Briefing cadence and reviewer criteria are not fully enumerated in the desk summary of the redline.
Sources
- Anthropic: “Anthropic’s Responsible Scaling Policy” (Version 3.2 effective April 29, 2026)
- RSP document and redline for Version 3.2.