Versions

One version is on record: the copy we retrieved on 26 Sep 2026. We know of no other.

  1. 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
    18 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. We know of only one version of it.

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. 4 of the 18 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 Grok 4.20 System Card
ModelEvaluationConditionValueLocationChecked
Grok 4Alignment auditVerbalized evaluation awareness rateacross all auditsTable 3Verified
Grok 4.20 (single-agent)Alignment auditVerbalized evaluation awareness ratesingle-agent; across all auditsTable 3Verified

M3 Sabotage and sandbagging

2 values

M3 Sabotage and sandbagging: values in the Grok 4.20 System Card
ModelEvaluationConditionValueLocationChecked
Grok 4.20 (single-agent)Alignment auditSubversive action rate vs usersingle-agent; unethical-operator scenariosTable 3Unverified
Grok 4.20 (single-agent)Alignment auditSubversive action rate vs xAIsingle-agentTable 3Unverified

M5 Honesty and hallucination

2 values

M5 Honesty and hallucination: values in the Grok 4.20 System Card
ModelEvaluationConditionValueLocationChecked
Grok 4.20 (single-agent)HLE calibrationRMS calibration errorsingle-agentTable 2Unverified
Grok 4.20 (single-agent)MASKDishonesty ratesingle-agentTable 2Unverified

M6 Sycophancy

2 values

M6 Sycophancy: values in the Grok 4.20 System Card
ModelEvaluationConditionValueLocationChecked
Grok 4.20 (single-agent)SycophancySycophancy ratesingle-agentTable 2Unverified
Grok 4.20 (single-agent)Alignment auditDelusion validation ratesingle-agentTable 3Unverified

M7 Harmful compliance and over-refusal

4 values

M7 Harmful compliance and over-refusal: values in the Grok 4.20 System Card
ModelEvaluationConditionValueLocationChecked
Grok 4.20 (single-agent)AgentHarmAnswer rate on harmful agentic taskssingle-agent; no attackTable 1Unverified
Grok 4.1Alignment auditCooperation with misusechatTable 3Verified
Grok 4.20 (single-agent)Alignment auditCooperation with misusesingle-agent; chatTable 3Unverified
Grok 4.20 (single-agent)Alignment auditMisuse with system-prompt overridesingle-agent; system prompt overrideTable 3Unverified

M8 Jailbreak robustness

1 value

M8 Jailbreak robustness: values in the Grok 4.20 System Card
ModelEvaluationConditionValueLocationChecked
Grok 4.20 (single-agent)JailbreaksAnswer rate on should-refuse promptssingle-agent; user jailbreakTable 1Unverified

M9 Prompt injection

1 value

M9 Prompt injection: values in the Grok 4.20 System Card
ModelEvaluationConditionValueLocationChecked
Grok 4.20 (single-agent)AgentDojoPrompt-injection attack success ratesingle-agentTable 1Unverified

M10 Dangerous capabilities and risk determinations

4 values

M10 Dangerous capabilities and risk determinations: values in the Grok 4.20 System Card
ModelEvaluationConditionValueLocationChecked
Grok 4.20Overall risk conclusionwith safeguardsConclusionUnverified
Grok 4.20 (single-agent)CloningScenariosAccuracywithout safeguards; single-agentTable 4Unverified
Grok 4.20 (single-agent)CybenchUnguided success ratewithout safeguards; single-agentTable 4Verified
Grok 4.20 (single-agent)VCTAccuracywithout safeguards; single-agentTable 4Unverified

Extraction coverage

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

Our note First xAI card with an alignment audit (Petri 2.0)

All 18 values were read in the document itself.

All 18 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).