Versions

Three versions are on record, oldest first. We hold a copy of one; the others are known only by their date.

  1. 5 Aug 2025

    First published version

    Date from
    our source registry
    Copy
    Known to exist; no copy held
  2. 15 Sep 2025

    Revision

    Date from
    the document's changelog
    Copy
    Known from the changelog; no copy held

    What changed CBRN partner acknowledgments

  3. 26 Sep 2026

    Date retrieved

    Copy retrieved 26 Sep 2026

    Date from
    the date we retrieved it; the copy states no version date
    Copy
    Copy retrieved on 26 Sep 2026
    Values
    34 values recorded from this version

No version has a file hash or an archived snapshot yet. From dataset v0.2 each retrieved version carries both (Methodology §7).

Revisions

No revisions are recorded for this document. Without copies of the earlier versions, changes between them are not recorded value by value; what we know of each version is listed below.

Values

Every value we recorded from this document, grouped by metric family and ordered by where the document prints it. Location is the section, table or page as the document numbers it. 1 of the 34 has been blind-verified: a second reader found the same value without seeing ours.

KeyVerifiedUnverifiedDisputedCorrected read off a figure or stated in wordsA value opens its source and history.

M2 Reward hacking

18 values

M2 Reward hacking: values in the Claude Opus 4.1 System Card (addendum)
ModelEvaluationConditionValueLocationChecked
Claude Opus 4Impossible tasksClassifier hack rateanti-hack promptTable 5.BUnverified
Claude Opus 4Impossible tasksClassifier hack rateno anti-hack promptTable 5.BUnverified
Claude Opus 4Reward-hack-prone coding tasksClassifier hack rateno anti-hack promptTable 5.BUnverified
Claude Opus 4Reward-hack-prone coding tasksHidden-test hack rateno anti-hack promptTable 5.BUnverified
Claude Opus 4.1Impossible tasksClassifier hack rateanti-hack promptTable 5.BUnverified
Claude Opus 4.1Impossible tasksClassifier hack rateno anti-hack promptTable 5.BUnverified
Claude Opus 4.1Reward-hack-prone coding tasksClassifier hack rateno anti-hack promptTable 5.BUnverified
Claude Opus 4.1Reward-hack-prone coding tasksHidden-test hack rateno anti-hack promptTable 5.BUnverified
Claude Opus 4.1Training distributionClassifier hack rateenvironment 1Table 5.BUnverified
Claude Opus 4.1Training distributionClassifier hack rateenvironment 2Table 5.BUnverified
Claude Sonnet 3.7Impossible tasksClassifier hack rateanti-hack promptTable 5.BUnverified
Claude Sonnet 3.7Impossible tasksClassifier hack rateno anti-hack promptTable 5.BUnverified
Claude Sonnet 3.7Reward-hack-prone coding tasksClassifier hack rateno anti-hack promptTable 5.BUnverified
Claude Sonnet 3.7Reward-hack-prone coding tasksHidden-test hack rateno anti-hack promptTable 5.BUnverified
Claude Sonnet 4Impossible tasksClassifier hack rateanti-hack promptTable 5.BUnverified
Claude Sonnet 4Impossible tasksClassifier hack rateno anti-hack promptTable 5.BUnverified
Claude Sonnet 4Reward-hack-prone coding tasksClassifier hack rateno anti-hack promptTable 5.BUnverified
Claude Sonnet 4Reward-hack-prone coding tasksHidden-test hack rateno anti-hack promptTable 5.BUnverified

M4 Misalignment audits

1 value

M4 Misalignment audits: values in the Claude Opus 4.1 System Card (addendum)
ModelEvaluationConditionValueLocationChecked
Claude Opus 4.1Automated behavioral auditReduction in cooperation with egregious misusevs Opus 4; share of transcripts scoring >5/10Sec 4.1 / Fig 4.1.AUnverified

M7 Harmful compliance and over-refusal

6 values

M7 Harmful compliance and over-refusal: values in the Claude Opus 4.1 System Card (addendum)
ModelEvaluationConditionValueLocationChecked
Claude Opus 4Single-turn violative requestsHarmless response rateoverall across thinking modesTable 2.1.AUnverified
Claude Opus 4.1Single-turn violative requestsHarmless response rateextended thinkingTable 2.1.AUnverified
Claude Opus 4.1Single-turn violative requestsHarmless response rateno extended thinkingTable 2.1.AUnverified
Claude Opus 4.1Single-turn violative requestsHarmless response rateoverall across thinking modesTable 2.1.AUnverified
Claude Opus 4Single-turn benign requestsOver-refusal rateoverall across thinking modesTable 2.1.BUnverified
Claude Opus 4.1Single-turn benign requestsOver-refusal rateoverall across thinking modesTable 2.1.BUnverified

M10 Dangerous capabilities and risk determinations

9 values

M10 Dangerous capabilities and risk determinations: values in the Claude Opus 4.1 System Card (addendum)
ModelEvaluationConditionValueLocationChecked
Claude Opus 4.1RSP deployment standardASL standard deployed underNone statedSec 6Unverified
Claude Opus 4Long-form virology task 1ScoreNone statedSec 6.3.1Unverified
Claude Opus 4.1Long-form virology task 1ScoreNone statedSec 6.3.1Unverified
Claude Opus 4AI R&D kernel taskBest speedupNone statedSec 6.4.1Verified
Claude Opus 4SWE-bench Verified (hard subset)Problems solved, pass@1 avgof 42 problemsSec 6.4.1Unverified
Claude Opus 4.1AI R&D kernel taskBest speedupNone statedSec 6.4.1Unverified
Claude Opus 4.1SWE-bench Verified (hard subset)Problems solved, pass@1 avgof 42 problemsSec 6.4.1Unverified
Claude Opus 4CybenchCTF challenges solvedof 35-challenge subsetSec 6.5.1Unverified
Claude Opus 4.1CybenchCTF challenges solvedof 35-challenge subsetSec 6.5.1Unverified

Extraction coverage

What we read of this document, and where each value was read.

Our note Read in full; carries corrected Opus 4 / Sonnet 4 reward-hacking numbers

All 34 values were read in the document itself.

All 34 values were extracted for version 0 of the dataset through a web reader, which did not always reach the later sections of long PDFs (Methodology §3).