Claim · 2025-05-22 · publication · Stated in source · Supported

Anthropic's ASL-3 activation post states that experiments showed access to Claude Sonnet 3.7 helped participants do somewhat better on tasks related to CBRN weapon acquisition than those with standard internet access, and that all participants' plans still had critical failures.

One recorded claim. The claim panel gives its qualification, the source passages with their place in each source, and how it was checked.

What it connects

Activating AI Safety Level 3 protections → states evaluation result → Anthropic states that experiments showed access to Claude Sonnet 3.7 helped participants do somewhat better on tasks related to CBRN weapon acquisition than those with standard internet access, while all participants' plans still had critical failures.

Where it appears

No collection page holds this claim. It is listed with its records in the explorer.

Claim and evidence

Activating AI Safety Level 3 protectionsstates evaluation result ↓Anthropic states that experiments showed access to Claude Sonnet 3.7 helped participants do somewhat better on tasks related to CBRN weapon acquisition than those with standard internet access, while all participants' plans still had critical failures.

Machine-checked, not reviewed by an editorStated in sourceSupported

Supporting evidence: primary · 1 source

Subject version: As published on anthropic.com on 22 May 2025 · Published 2025-05-22 · Effective date not established
Date: 2025-05-22 · publication
Evaluated system: Claude Sonnet 3.7
What the report measured: Whether access to the model improved participants' performance on tasks related to CBRN weapon acquisition, against a standard internet baseline
By what kind of method: Human-participant experiments comparing a model-access condition with a standard internet condition
The result, in the report’s words: Participants with model access did somewhat better than those with standard internet access
The caveats the report gives: The post states that all participants' plans still had critical failures, and gives the result as supporting evidence for an expectation about future models rather than as a full report.

Scope and limitsA comparative result as the developer reports it in a footnote, for the model and the participant tasks it describes, with the post's own statement that all plans still had critical failures. It is not generalised beyond those experiments and does not quantify any uplift. The tasks are referred to only at the generality of the post's own wording.

The evidence 01

primary sourceSupports the claim

Activating AI Safety Level 3 protections

Anthropic · Publication date unknown

Cite the original source

Cite the original document, not Selekoa. Check these details against it.

Source excerpt

“Experiments showed that access to Claude Sonnet 3.7 helped participants do somewhat better on tasks related to CBRN weapon acquisition than those with standard internet access (although all participants' plans still had critical failures).”

Location in source Footnote 2

How it was read Retrieved directly from the publisher

Machine-checked · 2026-10-30

Source details and check notes

Authors: Anthropic
Version: Retrieved for the candidate pack on 2026-10-28 (SHA-256 411f66885bacd8769dd6718e78f40eb0df80a4983dbde9d0d49f0323746e8656). Page labelled 'Published May 22, 2025', with six footnotes, as served on 2026-10-28. The accompanying full report (a separate PDF) was not retrieved. Retrieved again at admission on 2026-10-30 (SHA-256 35fa989dfc4f36842ef4f1dda2cfbdda9f9fc9f63524b310fd869ed2f32ab1e8). Its wording can change; earlier wording is not authenticated.
Canonical address: Open the canonical source ↗
Accessed: 2026-10-28
Language: English (en)
Attribution: Anthropic, “Activating AI Safety Level 3 protections”.
Reuse: No open redistribution licence asserted. Bibliographic metadata, editorial paraphrases and short attributed quotations (at most 50 words each). Recheck terms before wider redistribution.
Check notes: Candidate pack passage, AI and biosecurity (research/ai-bio). The quotation (34 words) matched the text retrieved for the candidate pack on 2026-10-28 and again at admission on 2026-10-30, in a fresh retrieval (SHA-256 35fa989dfc4f36842ef4f1dda2cfbdda9f9fc9f63524b310fd869ed2f32ab1e8; the bytes differ from the pack’s retrieval). Matching ignored differences of whitespace, case, quotation marks, dashes and Unicode compatibility forms, so the source’s typography can differ. A cross-family machine check (2026-10-30, item 31) re-read the record’s sources and found full support. No human editorial review.

Status and meaning

Editorial status. Not reviewed by an editor. An exploratory record that no editor has accepted. This status does not say whether a person has read any linked passage. Each evidence record above shows how it was checked.

Relation. A report states the result of an evaluation it ran or commissioned, in its own words, with what was measured, by what kind of method, and the caveats the report itself gives (schema 2.22). The record is the report’s statement, not a verified finding: it says nothing about another system, a later version or real-world risk, and a null or limiting result is not evidence that no effect exists.

Date. As stated in the post of 22 May 2025, reporting experiments it does not date.

Record ID ab-ev-anthropic-asl3-sonnet37-uplift