Safety Card Ledger

Claude Mythos Preview System Card

A system card by Anthropic about Claude Mythos Preview, published 7 Apr 2026. We recorded 43 values from it.

Developer
Anthropic
Type
System card
Model covered
Claude Mythos Preview
Published
7 Apr 2026
Original
www-cdn.anthropic.com
Archived copy
No archived copy yet
Changelog
Has a changelog

Data as of 26 Sep 2026 · Dataset v0.1 · Methodology v0.1 · Changelog

Versions

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

  1. 7 Apr 2026

    First published version

    Date from
    our source registry
    Copy
    Known to exist; no copy held

  2. 8 Apr 2026

    Revision

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

  3. 14 Apr 2026

    Revision

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

  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
    43 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 43 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

1 value

ModelEvaluationConditionValueLocationChecked
Claude Mythos PreviewEvaluation awareness in audit transcriptsShare of transcripts with eval awarenesswhite-box/unverbalized detectionSec 4.4-4.5Unverified

M2 Reward hacking

7 values

ModelEvaluationConditionValueLocationChecked
Claude Mythos PreviewGUI computer-use impossible tasksReward hack rateanti-hack promptSec 4.2Unverified
Claude Mythos PreviewImpossible tasksReward hack rateanti-hack promptSec 4.2Unverified
Claude Mythos PreviewImpossible tasksReward hack rateno anti-hack promptSec 4.2Unverified
Claude Opus 4.6GUI computer-use impossible tasksReward hack rateanti-hack promptSec 4.2Unverified
Claude Opus 4.6Impossible tasksReward hack rateanti-hack promptSec 4.2Unverified
Claude Opus 4.6Impossible tasksReward hack rateno anti-hack promptSec 4.2Unverified
Claude Sonnet 4.6GUI computer-use impossible tasksReward hack rateanti-hack promptSec 4.2Unverified

M3 Sabotage and sandbagging

4 values

ModelEvaluationConditionValueLocationChecked
Claude Mythos PreviewSandbagging transcript reviewPossible sandbagging cases20 flagged transcripts human-reviewedSec 4.4.2Unverified
Claude Mythos PreviewMinimal-LinuxBenchStealth success rateCoT visible to monitorSec 4.4.3Unverified
Claude Mythos PreviewSHADE-ArenaStealth success rateextended thinkingSec 4.4.3Unverified
Claude Mythos PreviewSHADE-ArenaStealth success ratehidden scratchpadSec 4.4.3Unverified

M4 Misalignment audits

4 values

ModelEvaluationConditionValueLocationChecked
Claude Mythos PreviewInternal deployment monitoringCovering tracks after disallowed actionsearlier snapshotsSec 4.1Unverified
Claude Mythos PreviewDestructive actions (synthetic backend)Destructive action rateNone statedSec 4.2Unverified
Claude Opus 4.6Destructive actions (synthetic backend)Destructive action rateNone statedSec 4.2Unverified
Claude Mythos PreviewAutomated behavioral auditOverall misaligned behavior scoreNone statedSec 4.2.3Unverified

M5 Honesty and hallucination

6 values

ModelEvaluationConditionValueLocationChecked
Claude Mythos Preview100Q-HardCorrect answer rateNone statedSec 4.3Unverified
Claude Mythos PreviewFalse-premise honestyHonest pushback rateNone statedSec 4.3Unverified
Claude Mythos PreviewMASKHonesty rate under pressureNone statedSec 4.3Unverified
Claude Opus 4.6100Q-HardCorrect answer rateNone statedSec 4.3Unverified
Claude Opus 4.6False-premise honestyHonest pushback rateNone statedSec 4.3Unverified
Claude Opus 4.6MASKHonesty rate under pressureNone statedSec 4.3Unverified

M6 Sycophancy

2 values

ModelEvaluationConditionValueLocationChecked
Claude Mythos PreviewAutomated behavioral auditSycophancy scoreNone statedSec 4.2.3Verified
Claude Opus 4.6Automated behavioral auditSycophancy scoreNone statedSec 4.2.3Unverified

M10 Dangerous capabilities and risk determinations

17 values

ModelEvaluationConditionValueLocationChecked
Claude Mythos PreviewRelease decisionDeployment scopeNone statedSec 1Unverified
Claude Mythos PreviewAutonomy threat model 2 (automated R&D)RSP determinationNone statedSec 2Unverified
Claude Mythos PreviewCB-1 threat modelRSP determinationNone statedSec 2Unverified
Claude Mythos PreviewCB-2 threat modelRSP determinationNone statedSec 2Unverified
Claude Mythos PreviewMultimodal virology (VCT)ScoreNone statedSec 2.2Unverified
Claude Mythos PreviewVirology protocol uplift trialMean critical failuresmodel-assistedSec 2.2Unverified
Claude Opus 4.6Multimodal virology (VCT)ScoreNone statedSec 2.2Unverified
Claude Mythos PreviewInternal AI R&D surveyRespondents saying drop-in entry-level researcher nowof 18 respondentsSec 2.3Unverified
Claude Mythos PreviewAI R&D kernel taskBest speedupNone statedSec 2.3 tableUnverified
Claude Mythos PreviewAI R&D LLM training taskAverage speedupNone statedSec 2.3 tableVerified
Claude Mythos PreviewAI R&D novel compiler taskPass rateNone statedSec 2.3 tableUnverified
Claude Mythos PreviewInternal AI R&D suite 2Suite scoreNone statedSec 2.3 tableUnverified
Claude Opus 4.6AI R&D kernel taskBest speedupNone statedSec 2.3 tableUnverified
Claude Opus 4.6AI R&D LLM training taskAverage speedupNone statedSec 2.3 tableUnverified
Claude Mythos PreviewCybenchSuccess rate, pass@135 challengesSec 3Unverified
Claude Mythos PreviewCyberGymTargeted vuln reproduction rateNone statedSec 3.3.2Unverified
Claude Opus 4.6CyberGymTargeted vuln reproduction rateNone statedSec 3.3.2Unverified

M11 Self-preservation

1 value

ModelEvaluationConditionValueLocationChecked
Claude Mythos PreviewSandbox escape incidentReal-world incident reportedinternal useSec 4.1Unverified

M12 Chain-of-thought monitorability

1 value

ModelEvaluationConditionValueLocationChecked
Claude Mythos PreviewEncoded reasoning (scratchpad paraphrase)Pass@1 with paraphrased scratchpadNone statedSec 4.4.1Unverified

Extraction coverage

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

Our note Web reader stopped ~p.49; limited-release model; first card under RSP v3.0

Of the 43 values, 15 were read in the document itself, 2 on a developer summary page that restates it and 26 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).

Developer summary pages used

Of the 43 values, 3 were read off a figure; they are marked * and left out of charts by default.

All 43 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).