Values by metric family

Each value as the document printed it. Select a value for its source, its checks and its history. Values from other documents are listed after the model’s own, each marked as the first report of that measure or as a restatement.

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

M1 · Evaluation awareness 3 values

M1 Evaluation awareness: values about Claude Sonnet 4.5
Evaluation and metricValueDocument
From its own documentClaude Sonnet 4.5 System Card · 29 Sep 2025
Automated behavioral auditShare of transcripts verbalizing eval awarenessClaude Sonnet 4.5 System Card29 Sep 2025
Evaluation awarenessUnambiguous references to being evaluatedRun by Apollo ResearchClaude Sonnet 4.5 System Card29 Sep 2025
Evaluation awarenessUnprompted verbalized eval awareness rateRun by UK AI Security InstituteClaude Sonnet 4.5 System Card29 Sep 2025

M2 · Reward hacking 6 values

M2 Reward hacking: values about Claude Sonnet 4.5
Evaluation and metricValueDocument
From its own documentClaude Sonnet 4.5 System Card · 29 Sep 2025
Impossible tasksClassifier hack rateCondition: anti-hack promptClaude Sonnet 4.5 System Card29 Sep 2025
Impossible tasksClassifier hack rateCondition: no anti-hack promptRestated later: compare with the later valueClaude Sonnet 4.5 System Card29 Sep 2025Restated later: compare with the later value
Reward-hack-prone coding tasksClassifier hack rateCondition: no anti-hack promptClaude Sonnet 4.5 System Card29 Sep 2025
Reward-hack-prone coding tasksHidden-test hack rateCondition: no anti-hack promptClaude Sonnet 4.5 System Card29 Sep 2025
Training distributionClassifier hack rateClaude Sonnet 4.5 System Card29 Sep 2025
From other documents
Impossible tasksClassifier hack rateCondition: no anti-hack promptClaude Haiku 4.5 System Card15 Oct 2025Restated: compare with the earlier valueClaude Haiku 4.5 System Card15 Oct 2025Restated: compare with the earlier value

M3 · Sabotage and sandbagging 1 value

M3 Sabotage and sandbagging: values about Claude Sonnet 4.5
Evaluation and metricValueDocument
From its own documentClaude Sonnet 4.5 System Card · 29 Sep 2025
SHADE-ArenaStealth success rateCondition: no extended thinkingClaude Sonnet 4.5 System Card29 Sep 2025

M4 · Misalignment audits 1 value

M4 Misalignment audits: values about Claude Sonnet 4.5
Evaluation and metricValueDocument
From its own documentClaude Sonnet 4.5 System Card · 29 Sep 2025
Automated behavioral auditImprovement on overall misaligned-behavior scoreCondition: vs Sonnet 4Claude Sonnet 4.5 System Card29 Sep 2025

M6 · Sycophancy 1 value

M6 Sycophancy: values about Claude Sonnet 4.5
Evaluation and metricValueDocument
From other documents
Sycophancy in prefilled conversationsCourse-correction rateCondition: continuing sycophantic conversationClaude Opus 4.5 System Card24 Nov 2025First reportedClaude Opus 4.5 System Card24 Nov 2025First reported

M7 · Harmful compliance and over-refusal 22 values

M7 Harmful compliance and over-refusal: values about Claude Sonnet 4.5
Evaluation and metricValueDocument
From its own documentClaude Sonnet 4.5 System Card · 29 Sep 2025
Malicious agentic codingSafety scoreCondition: without safeguardsRestated later: compare with the later valueClaude Sonnet 4.5 System Card29 Sep 2025Restated later: compare with the later value
Malicious Claude Code useRefusal rate, covert maliciousCondition: with mitigationsClaude Sonnet 4.5 System Card29 Sep 2025
Malicious Claude Code useRefusal rate, covert maliciousCondition: without safeguardsClaude Sonnet 4.5 System Card29 Sep 2025
Malicious Claude Code useRefusal rate, overt maliciousCondition: without safeguardsClaude Sonnet 4.5 System Card29 Sep 2025
Malicious Claude Code useSuccess rate, dual-use requestsCondition: without safeguardsClaude Sonnet 4.5 System Card29 Sep 2025
Single-turn benign requestsOver-refusal rateCondition: overall across thinking modesRestated later: compare with the later valueClaude Sonnet 4.5 System Card29 Sep 2025Restated later: compare with the later value
Single-turn violative requestsHarmless response rateCondition: extended thinkingClaude Sonnet 4.5 System Card29 Sep 2025
Single-turn violative requestsHarmless response rateCondition: no extended thinkingRestated later: compare with the later valueClaude Sonnet 4.5 System Card29 Sep 2025Restated later: compare with the later value
Single-turn violative requestsHarmless response rateCondition: overall across thinking modesRestated later: compare with the later valueClaude Sonnet 4.5 System Card29 Sep 2025Restated later: compare with the later value
From other documents
Malicious agentic codingSafety scoreCondition: without safeguardsClaude Haiku 4.5 System Card15 Oct 2025Restated: compare with the earlier valueClaude Haiku 4.5 System Card15 Oct 2025Restated: compare with the earlier value
Malicious Claude Code useRefusal rate, malicious requestsCondition: without safeguardsClaude Haiku 4.5 System Card15 Oct 2025First reportedRestated later: compare with the later valueClaude Haiku 4.5 System Card15 Oct 2025First reportedRestated later: compare with the later value
Malicious Claude Code useSuccess rate, dual-use & benignCondition: without safeguardsClaude Haiku 4.5 System Card15 Oct 2025First reportedClaude Haiku 4.5 System Card15 Oct 2025First reported
Single-turn violative requestsHarmless response rateCondition: no extended thinkingClaude Haiku 4.5 System Card15 Oct 2025Restated: compare with the earlier valueClaude Haiku 4.5 System Card15 Oct 2025Restated: compare with the earlier value
Malicious agentic codingSafety scoreCondition: without safeguardsClaude Opus 4.5 System Card24 Nov 2025Restated: compare with the earlier valueClaude Opus 4.5 System Card24 Nov 2025Restated: compare with the earlier value
Malicious Claude Code useRefusal rate, malicious requestsCondition: without safeguardsClaude Opus 4.5 System Card24 Nov 2025Restated: compare with the earlier valueClaude Opus 4.5 System Card24 Nov 2025Restated: compare with the earlier value
Malicious computer useRefusal rateCondition: without safeguardsClaude Opus 4.5 System Card24 Nov 2025First reportedClaude Opus 4.5 System Card24 Nov 2025First reported
Single-turn benign requestsOver-refusal rateCondition: overall across thinking modes; all languagesClaude Opus 4.5 System Card24 Nov 2025Restated: compare with the earlier valueClaude Opus 4.5 System Card24 Nov 2025Restated: compare with the earlier value
Single-turn violative requestsHarmless response rateCondition: overall across thinking modes; all languagesClaude Opus 4.5 System Card24 Nov 2025Restated: compare with the earlier valueClaude Opus 4.5 System Card24 Nov 2025Restated: compare with the earlier value
Higher-difficulty benign requestsOver-refusal rateCondition: overall across thinking modesClaude Opus 4.6 System CardFeb 2026First reportedClaude Opus 4.6 System CardFeb 2026First reported
Single-turn benign requestsOver-refusal rateCondition: overall across thinking modesClaude Opus 4.6 System CardFeb 2026Restated: compare with the earlier valueClaude Opus 4.6 System CardFeb 2026Restated: compare with the earlier value
Single-turn violative requestsHarmless response rateCondition: overall across thinking modesClaude Opus 4.6 System CardFeb 2026Restated: compare with the earlier valueClaude Opus 4.6 System CardFeb 2026Restated: compare with the earlier value
Single-turn violative requestsHarmless response rateCondition: overall across thinking modesClaude Sonnet 4.6 System Card17 Feb 2026Restated: compare with the earlier valueClaude Sonnet 4.6 System Card17 Feb 2026Restated: compare with the earlier value

M9 · Prompt injection 13 values

M9 Prompt injection: values about Claude Sonnet 4.5
Evaluation and metricValueDocument
From its own documentClaude Sonnet 4.5 System Card · 29 Sep 2025
Computer-use prompt injectionAttack prevention rateCondition: with safeguardsClaude Sonnet 4.5 System Card29 Sep 2025
Computer-use prompt injectionAttack prevention rateCondition: without safeguardsClaude Sonnet 4.5 System Card29 Sep 2025
MCP prompt injectionAttack prevention rateCondition: with safeguardsClaude Sonnet 4.5 System Card29 Sep 2025
MCP prompt injectionAttack prevention rateCondition: without safeguardsClaude Sonnet 4.5 System Card29 Sep 2025
Tool-use prompt injectionAttack prevention rateCondition: with safeguardsClaude Sonnet 4.5 System Card29 Sep 2025
Tool-use prompt injectionAttack prevention rateCondition: without safeguardsClaude Sonnet 4.5 System Card29 Sep 2025
From other documents
Coding prompt injection (adaptive attacker)Attack success rateCondition: extended thinking; without safeguards; 1 attemptClaude Opus 4.5 System Card24 Nov 2025First reportedClaude Opus 4.5 System Card24 Nov 2025First reported
Coding prompt injection (adaptive attacker)Attack success rateCondition: extended thinking; without safeguards; 200 attemptsClaude Opus 4.5 System Card24 Nov 2025First reportedClaude Opus 4.5 System Card24 Nov 2025First reported
Coding prompt injection (adaptive attacker)Attack success rateCondition: no extended thinking; without safeguards; 1 attemptClaude Opus 4.5 System Card24 Nov 2025First reportedClaude Opus 4.5 System Card24 Nov 2025First reported
Coding prompt injection (adaptive attacker)Attack success rateCondition: no extended thinking; without safeguards; 200 attemptsClaude Opus 4.5 System Card24 Nov 2025First reportedClaude Opus 4.5 System Card24 Nov 2025First reported
Computer use prompt injection (adaptive attacker)Attack success rateCondition: extended thinking; with safeguards; 200 attemptsClaude Opus 4.5 System Card24 Nov 2025First reportedClaude Opus 4.5 System Card24 Nov 2025First reported
Computer use prompt injection (adaptive attacker)Attack success rateCondition: extended thinking; without safeguards; 1 attemptClaude Opus 4.5 System Card24 Nov 2025First reportedClaude Opus 4.5 System Card24 Nov 2025First reported
Computer use prompt injection (adaptive attacker)Attack success rateCondition: extended thinking; without safeguards; 200 attemptsClaude Opus 4.5 System Card24 Nov 2025First reportedClaude Opus 4.5 System Card24 Nov 2025First reported

M10 · Dangerous capabilities and risk determinations 2 values

M10 Dangerous capabilities and risk determinations: values about Claude Sonnet 4.5
Evaluation and metricValueDocument
From its own documentClaude Sonnet 4.5 System Card · 29 Sep 2025
CybenchSuccess rate, pass@30Condition: 37 of 40 challengesClaude Sonnet 4.5 System Card29 Sep 2025
RSP deployment standardASL standard deployed underClaude Sonnet 4.5 System Card29 Sep 2025

Restated in later documents

A later document reported values about Claude Sonnet 4.5 again. Each pair is shown side by side: both values are kept with their own documents, and the later one does not replace the earlier.

Values about Claude Sonnet 4.5 restated in later documents
Earlier valueLater value
M2 · Reward hacking
Impossible tasksClassifier hack rateCondition: no anti-hack prompt
Claude Sonnet 4.5 System Card29 Sep 2025 · its own documentClaude Haiku 4.5 System Card15 Oct 2025No reason stated
M7 · Harmful compliance and over-refusal
Malicious agentic codingSafety scoreCondition: without safeguards
Claude Sonnet 4.5 System Card29 Sep 2025 · its own documentClaude Haiku 4.5 System Card15 Oct 2025No reason stated
Malicious agentic codingSafety scoreCondition: without safeguards
Claude Sonnet 4.5 System Card29 Sep 2025 · its own documentClaude Opus 4.5 System Card24 Nov 2025Reason stated: metric renamed to refusal rate
Single-turn benign requestsOver-refusal rateCondition: overall across thinking modesLater condition: overall across thinking modes; all languages
Claude Sonnet 4.5 System Card29 Sep 2025 · its own documentClaude Opus 4.5 System Card24 Nov 2025Reason stated: all-language totals instead of English only
Single-turn benign requestsOver-refusal rateCondition: overall across thinking modes
Claude Sonnet 4.5 System Card29 Sep 2025 · its own documentClaude Opus 4.6 System CardFeb 2026No reason stated
Single-turn violative requestsHarmless response rateCondition: no extended thinking
Claude Sonnet 4.5 System Card29 Sep 2025 · its own documentClaude Haiku 4.5 System Card15 Oct 2025No reason stated
Single-turn violative requestsHarmless response rateCondition: overall across thinking modesLater condition: overall across thinking modes; all languages
Claude Sonnet 4.5 System Card29 Sep 2025 · its own documentClaude Opus 4.5 System Card24 Nov 2025Reason stated: all-language totals instead of English only
Single-turn violative requestsHarmless response rateCondition: overall across thinking modes
Claude Sonnet 4.5 System Card29 Sep 2025 · its own documentClaude Opus 4.6 System CardFeb 2026No reason stated
Single-turn violative requestsHarmless response rateCondition: overall across thinking modes
Claude Sonnet 4.5 System Card29 Sep 2025 · its own documentClaude Sonnet 4.6 System Card17 Feb 2026No reason stated
Malicious Claude Code useRefusal rate, malicious requestsCondition: without safeguards
Claude Haiku 4.5 System Card15 Oct 2025 · first reportedClaude Opus 4.5 System Card24 Nov 2025No reason stated

Risk determinations

The developer's formal decisions about Claude Sonnet 4.5 under its framework, as printed. Levels from different frameworks do not map onto one another.

Risk determinations about Claude Sonnet 4.5
DomainLevel as printedFramework and document
OverallASL-3 Responsible Scaling Policy Claude Sonnet 4.5 System Card · 29 Sep 2025Responsible Scaling Policy Claude Sonnet 4.5 System Card · 29 Sep 2025