Benchmark Almanac

Malicious use of Claude Code (Anthropic)

Claude Mythos Preview 96.7% leads since Apr 2026 Only in system cards

Agents in Claude Code receive malicious and dual-use coding requests. Score is the refusal rate on malicious requests.

In-house Anthropic evaluation. The tasks are not public.

Made by Anthropic Released Sep 2025 Refusing harmful requests Tasks kept private Last in a system card Oct 2026 Last model added Oct 2026
Level
Expert Specialists with years of training, often over hours of work
Scored
Graded by a model A language model scores each response, usually against a rubric.
Answers
Agent task Many steps in a terminal, browser, desktop, app, game or robot.
16
Models scored in lab cards
18
System cards report it, from 1 lab

Claude Mythos Preview leads with 96.7%

Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.

16 models. Scroll sideways for the rest.

  1. 96.7 Claude Mythos Preview
  2. 95.1 Claude Opus 4.8
  3. 92.4 Claude Sonnet 5
  4. 91.2 Claude Opus 4.7
  5. 90.3 Claude Fable 5.1
  6. 90.3 Claude Fable 5
  7. 90.0 Claude Sonnet 4
  8. 89.0 Claude Opus 5
  9. 88.0 Claude Opus 4
  10. 85.2 Claude Sonnet 5.5
  11. 84.3 Claude Haiku 5.5
  12. 83.2 Claude Opus 4.6
  13. 79.8 Claude Opus 5.5
  14. 79.3 Claude Sonnet 4.6
  15. 77.8 Claude Opus 4.5
  16. 69.4 Claude Haiku 4.5
  1. 1 Claude Mythos Preview 96.7%
  2. 2 Claude Opus 4.8 95.1%
  3. 3 Claude Sonnet 5 92.4%
  4. 4 Claude Opus 4.7 91.2%
  5. 5 Claude Fable 5.1 90.3%
  6. 6 Claude Fable 5 90.3%
  7. 7 Claude Sonnet 4 90%
  8. 8 Claude Opus 5 89%
  9. 9 Claude Opus 4 88%
  10. 10 Claude Sonnet 5.5 85.2%
  11. 11 Claude Haiku 5.5 84.3%
  12. 12 Claude Opus 4.6 83.2%
  13. 13 Claude Opus 5.5 79.8%
  14. 14 Claude Sonnet 4.6 79.3%
  15. 15 Claude Opus 4.5 77.8%
  16. 16 Claude Haiku 4.5 69.4%
#ModelScoreSettingDateSource
1 Claude Mythos Preview 96.7% Apr 2026 Model card
2 Claude Opus 4.8 95.1% May 2026 Model card
3 Claude Sonnet 5 92.4% Jun 2026 Model card
4 Claude Opus 4.7 91.2% Apr 2026 Model card
5 Claude Fable 5.1 90.3% Sep 2026 Model card
6 Claude Fable 5 90.3% Jun 2026 Model card
7 Claude Sonnet 4 90% May 2025 Model card
8 Claude Opus 5 89% Jul 2026 Model card
9 Claude Opus 4 88% May 2025 Model card
10 Claude Sonnet 5.5 85.2% Sep 2026 Model card
11 Claude Haiku 5.5 84.3% Oct 2026 Model card
12 Claude Opus 4.6 83.2% Feb 2026 Model card
13 Claude Opus 5.5 79.8% Sep 2026 Model card
14 Claude Sonnet 4.6 79.3% Feb 2026 Model card
15 Claude Opus 4.5 77.8% Nov 2025 Model card
16 Claude Haiku 4.5 69.4% Oct 2025 Model card

2 models have held number one since its release in Sep 2025

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic
Lab claims Passed 85% in Sep 2025, by Claude Sonnet 4
Claude Mythos Preview 96.7%

The best score passed 85% in May 2025

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Claude Mythos Preview Claude Sonnet 4 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Claude Mythos Preview Claude Sonnet 4

Labs report it only in system cards

We read the headline results of 479 launch posts from 51 labs, and none of them list this benchmark. Some system cards do, below.

18 system cards report it, from 1 lab

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
7 Oct 2026 Claude Haiku 5.5 system card 84.3 Malicious use of Claude Code, refusal rate on malicious requests, Claude Code system prompt, no extra safeguards. Dual-use and benign success 98.9 (Table 5.1.1.A) Safety
28 Sep 2026 Claude Sonnet 5.5 system card 85.2 Malicious use of Claude Code, refusal rate on malicious requests with the Claude Code system prompt and no extra safeguards, dual-use success 98.4% (Table 5.1.1.A) Safety
22 Sep 2026 Claude Opus 5.5 system card 79.8 Malicious use of Claude Code, refusal rate on 61 malicious prompts run 10 times, Claude Code system prompt, no additional safeguards (dual-use and benign success 99.8%) (Table 5.1.1.A) Safety
1 Sep 2026 Claude Fable 5.1 system card 90.3 Malicious use of Claude Code, refusal rate on 61 malicious prompts x10, Claude Code system prompt, no additional safeguards, row labeled Mythos 5.1; dual-use and benign success rate 98.4% (Table 5.1.1.A) Safety
24 Jul 2026 Claude Opus 5 system card 89 Malicious use of Claude Code, refusal rate on 61 malicious prompts run 10 times, Claude Code system prompt, no additional safeguards (dual-use and benign success rate 99.82%) (Table 5.1.1.A) Safety
30 Jun 2026 Claude Sonnet 5 system card 92.4 Malicious use of Claude Code, refusal rate on 61 malicious prompts run 10 times, Claude Code system prompt, no additional safeguards (dual-use and benign success 91.55%) (Table 5.1.1.A) Safety
9 Jun 2026 Claude Fable 5 system card 90.3 Malicious use of Claude Code, refusal rate on 61 malicious prompts x10, Claude Code system prompt, no additional safeguards, row labeled Mythos 5 (section 5 reports Mythos 5 only and says Fable 5 inherits these behaviors); dual-use and benign success rate 99.64% (Table 5.1.1.A) Safety
28 May 2026 Claude Opus 4.8 system card 95.1 Malicious use of Claude Code, refusal rate on 61 malicious prompts run 10 times, Claude Code system prompt only with the FileRead reminder removed. Dual-use and benign success rate 92.12% (Table 5.1.1.A) Safety
16 Apr 2026 Claude Opus 4.7 system card 91.2 Malicious use of Claude Code, refusal rate on 61 malicious prompts run 10 times, Claude Code system prompt, without FileTool reminder. Dual-use and benign success rate 91.83 (Table 5.1.1.A) Safety
7 Apr 2026 Claude Mythos Preview system card 96.7 Malicious use of Claude Code, refusal rate on malicious requests with the Claude Code system prompt, no FileTool reminder, 12 harder cases added; dual-use and benign success 92.75 (Table 8.3.1.1.A) Safety
17 Feb 2026 Claude Sonnet 4.6 system card 79.3 Malicious use of Claude Code, refusal rate on malicious prompts without mitigations; 99.39 with mitigations, dual-use success 88.52 and 91.78 (Tables 5.1.2.A and 5.1.2.B) Safety
5 Feb 2026 Claude Opus 4.6 system card 83.2 Malicious use of Claude Code, refusal rate on 49 malicious prompts, without mitigations (Table 5.1.2.A); 99.59 with mitigations, dual-use and benign success 91.75 without and 95.59 with mitigations (Table 5.1.2.B) Safety
24 Nov 2025 Claude Opus 4.5 system card 77.8 Malicious use of Claude Code, refusal rate on 49 malicious prompts, without mitigations (Table 5.1.2.A; 97.35% with mitigations, dual-use and benign success 93.07% without and 96.52% with mitigations) Safety
15 Oct 2025 Claude Haiku 4.5 system card 69.4 Malicious use of Claude Code, 49 malicious prompts, refusal rate without mitigations; 96.33 with previous and 99.17 with new mitigations, dual-use and benign success 88.85 (Tables 3.1.2.A and 3.1.2.B) Safety
29 Sep 2025 Claude Sonnet 4.5 system card Malicious use of Claude Code, no combined malicious refusal rate. Without mitigations: overt refusal 95.56%, covert refusal 52.42%, dual-use success 87.00%. With mitigations: 100%, 96.31%, 91.64% (Tables 4.1.A and 4.1.B) Safety
5 Aug 2025 Claude Opus 4.1 system card addendum Malicious use of agentic coding, three evaluations, no number printed. Text says compliance similar to Claude Opus 4 (3.3) Safety
22 May 2025 Claude Sonnet 4 system card 90 Agentic coding malicious use, safety score without safeguards averaged across three evaluations (150 prohibited, 2 x 50 borderline problems) 90% (Table 3.3.A) Safety
22 May 2025 Claude Opus 4 system card 88 Agentic coding malicious use, safety score without safeguards averaged across three evaluations (150 prohibited, 2 x 50 borderline problems) 88% (Table 3.3.A) Safety

More in Safety and alignment

All 444 in Safety and alignment
  • FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
  • StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is. Inkling 98.6%
  • AgentDojo In use Agents do tool-use tasks while tools and data carry prompt injections. Attack success rate counts hijacked runs. Grok 4 Fast 0%
  • MASK In use Measures whether a model contradicts its own elicited beliefs under pressure to lie. Honesty separated from accuracy. Muse Spark Open 98.1%
  • Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts. GPT-6 Astra 8.50%
  • Agentic misalignment suite (Anthropic) In use Agents in simulated company settings face threats or goal conflicts that invite blackmail, leaking or sabotage. Score is the misaligned action rate. Claude Sonnet 4.5 0%