Versions

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

  1. 22 May 2025

    First published version

    Date from
    our source registry
    Copy
    Known to exist; no copy held
  2. 16 Jul 2025

    Revision

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

    What changed footnote and formatting

  3. 2 Sep 2025

    Revision

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

    What changed corrected Claude Code Impossible Tasks numbers

  4. 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
    45 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. 2 of the 45 have 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.

M1 Evaluation awareness

2 values

M1 Evaluation awareness: values in the Claude Opus 4 & Sonnet 4 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4Automated behavioral auditTranscripts flagged for situational awarenessof 414 audit transcriptsSec 4.1.2.2Verified
Claude Sonnet 3.7Automated behavioral auditTranscripts flagged for situational awarenessof 414 audit transcriptsSec 4.1.2.2Unverified

M2 Reward hacking

2 values

M2 Reward hacking: values in the Claude Opus 4 & Sonnet 4 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4Reward hacking (hard-coding)Avg decrease in hard-coding vs Sonnet 3.7vs Sonnet 3.7Sec 6Unverified
Claude Sonnet 4Reward hacking (hard-coding)Avg decrease in hard-coding vs Sonnet 3.7vs Sonnet 3.7Sec 6Unverified

M3 Sabotage and sandbagging

4 values

M3 Sabotage and sandbagging: values in the Claude Opus 4 & Sonnet 4 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4Subtle sabotage (agentic)Side-task success without monitor detectionhidden scratchpadTable 4.1.10.BUnverified
Claude Opus 4Subtle sabotage (agentic)Side-task success without monitor detectionscratchpad visible to monitorTable 4.1.10.BUnverified
Claude Sonnet 3.7Subtle sabotage (agentic)Side-task success without monitor detectionhidden scratchpadTable 4.1.10.BUnverified
Claude Sonnet 3.7Subtle sabotage (agentic)Side-task success without monitor detectionscratchpad visible to monitorTable 4.1.10.BUnverified

M4 Misalignment audits

4 values

M4 Misalignment audits: values in the Claude Opus 4 & Sonnet 4 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4Opportunistic blackmail scenarioShare of rollouts with blackmailreplacement model shares valuesSec 4.1.1.2Unverified
Claude Opus 4Automated behavioral auditSurprising high-impact actions foundaudit transcriptsSec 4.1.9Unverified
Claude Sonnet 3.7Automated behavioral auditSurprising high-impact actions foundaudit transcriptsSec 4.1.9Unverified
Claude Sonnet 4Automated behavioral auditSurprising high-impact actions foundaudit transcriptsSec 4.1.9Verified

M5 Honesty and hallucination

4 values

M5 Honesty and hallucination: values in the Claude Opus 4 & Sonnet 4 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4Reasoning-behavior monitorKnowingly hallucinated informationRL-style transcriptsTable 4.1.6.AUnverified
Claude Opus 4Reasoning-behavior monitorOutputs showing deception (overall)RL-style transcriptsTable 4.1.6.AUnverified
Claude Sonnet 3.7Reasoning-behavior monitorKnowingly hallucinated informationRL-style transcriptsTable 4.1.6.AUnverified
Claude Sonnet 3.7Reasoning-behavior monitorOutputs showing deception (overall)RL-style transcriptsTable 4.1.6.AUnverified

M6 Sycophancy

2 values

M6 Sycophancy: values in the Claude Opus 4 & Sonnet 4 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4Reasoning-behavior monitorOutputs showing sycophancyRL-style transcriptsTable 4.1.6.AUnverified
Claude Sonnet 3.7Reasoning-behavior monitorOutputs showing sycophancyRL-style transcriptsTable 4.1.6.AUnverified

M7 Harmful compliance and over-refusal

12 values

M7 Harmful compliance and over-refusal: values in the Claude Opus 4 & Sonnet 4 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4Single-turn violative requestsHarmless response rateextended thinkingTable 2.1.AUnverified
Claude Opus 4Single-turn violative requestsHarmless response rateno extended thinkingTable 2.1.AUnverified
Claude Opus 4Single-turn violative requestsHarmless response rateoverall across thinking modesTable 2.1.AUnverified
Claude Opus 4Single-turn violative requestsHarmless response rateoverall across thinking modes; with ASL-3 safeguardsTable 2.1.AUnverified
Claude Sonnet 3.7Single-turn violative requestsHarmless response rateoverall across thinking modesTable 2.1.AUnverified
Claude Sonnet 4Single-turn violative requestsHarmless response rateoverall across thinking modesTable 2.1.AUnverified
Claude Opus 4Single-turn benign requestsOver-refusal rateoverall across thinking modesTable 2.2.AUnverified
Claude Sonnet 3.7Single-turn benign requestsOver-refusal rateoverall across thinking modesTable 2.2.AUnverified
Claude Sonnet 4Single-turn benign requestsOver-refusal rateoverall across thinking modesTable 2.2.AUnverified
Claude Opus 4Malicious agentic codingSafety scorewithout safeguardsTable 3.3.AUnverified
Claude Sonnet 3.7Malicious agentic codingSafety scorewithout safeguardsTable 3.3.AUnverified
Claude Sonnet 4Malicious agentic codingSafety scorewithout safeguardsTable 3.3.AUnverified

M8 Jailbreak robustness

6 values

M8 Jailbreak robustness: values in the Claude Opus 4 & Sonnet 4 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4StrongREJECTBest-jailbreak StrongREJECT scoreextended thinkingTable 2.7.AUnverified
Claude Opus 4StrongREJECTBest-jailbreak StrongREJECT scoreno extended thinkingTable 2.7.AUnverified
Claude Sonnet 3.7StrongREJECTBest-jailbreak StrongREJECT scoreextended thinkingTable 2.7.AUnverified
Claude Sonnet 3.7StrongREJECTBest-jailbreak StrongREJECT scoreno extended thinkingTable 2.7.AUnverified
Claude Sonnet 4StrongREJECTBest-jailbreak StrongREJECT scoreextended thinkingTable 2.7.AUnverified
Claude Sonnet 4StrongREJECTBest-jailbreak StrongREJECT scoreno extended thinkingTable 2.7.AUnverified

M9 Prompt injection

6 values

M9 Prompt injection: values in the Claude Opus 4 & Sonnet 4 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4Computer-use prompt injectionAttack prevention ratewith safeguardsTable 3.2.AUnverified
Claude Opus 4Computer-use prompt injectionAttack prevention ratewithout safeguardsTable 3.2.AUnverified
Claude Sonnet 3.7Computer-use prompt injectionAttack prevention ratewith safeguardsTable 3.2.AUnverified
Claude Sonnet 3.7Computer-use prompt injectionAttack prevention ratewithout safeguardsTable 3.2.AUnverified
Claude Sonnet 4Computer-use prompt injectionAttack prevention ratewith safeguardsTable 3.2.AUnverified
Claude Sonnet 4Computer-use prompt injectionAttack prevention ratewithout safeguardsTable 3.2.AUnverified

M10 Dangerous capabilities and risk determinations

3 values

M10 Dangerous capabilities and risk determinations: values in the Claude Opus 4 & Sonnet 4 System Card
ModelEvaluationConditionValueLocationChecked
Claude Opus 4RSP deployment standardASL standard deployed underNone statedSec 1 / Sec 7Unverified
Claude Sonnet 4RSP deployment standardASL standard deployed underNone statedSec 1 / Sec 7Unverified
Claude Opus 4CybenchCTF challenges solvedof 39 challengesSec 7.4Unverified

Extraction coverage

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

Our note Web reader stopped at §4.2.1; later sections partly from secondary write-ups

Of the 45 values, 42 were read in the document itself and 3 in independent write-ups that quote it.

Values read somewhere other than the document stand in where the document's own section could not be read directly, and are flagged on every value (Methodology §2.2).

All 45 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).