Agentic misalignment suite (Anthropic)
Claude Sonnet 4.5 0% leads since Sep 2025 In use
Agents in simulated company settings face threats or goal conflicts that invite blackmail, leaking or sabotage. Score is the misaligned action rate.
Based on Anthropic's Agentic Misalignment research. The cards use internal variants.
Made by Anthropic Released Jun 2025 Deception and alignment All tasks public Last in a launch post 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.
7
Models scored in lab cards
1
Launch post lists it, from 1 lab
9
System cards report it, from 2 labs
4 models share the top score of 0%
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.
- 1 Claude Sonnet 4.5 0%
- 2
Muse Wonder Full 0% - 3 Muse Spark Open 0%
- 4 Muse Spark 1.3 0%
- 5 Muse Spark 1.1 1.10%
- 6 Muse Spark 47.7%
- 7 Claude Opus 4 84%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 0% | Sep 2025 | Model card | |
| 2 | Muse Wonder Full | 0% | Oct 2026 | Model card | |
| 3 | Muse Spark Open | 0% | Oct 2026 | Model card | |
| 4 | Muse Spark 1.3 | 0% | Oct 2026 | Model card | |
| 5 | Muse Spark 1.1 | 1.10% | Jul 2026 | Model card | |
| 6 | Muse Spark | 47.7% | Jun 2026 | Model card | |
| 7 | Claude Opus 4 | 84% | May 2025 | Model card |
2 models have held number one since its release in Jun 2025
Lab claims count from the day the card came out, independent runs from the model’s release.
Anthropic
Scores over time
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
One lab put it in a launch post in the last six months
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| 15 Oct 2026 | Muse Spark Open | 0 | Agentic Misalignment benchmark, misbehavior rate 0% (text, lower is better) |
9 system cards report it, from 2 labs
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.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| 23 Oct 2026 | Muse Spark 1.3 system card | 0 | Agentic Misalignment (Lynch et al.) harmful-action rate, Muse Spark 1.3 (API) (Table 2, Table 15) | Alignment |
| 12 Oct 2026 | Muse Wonder Full model card | 0 | Agentic Misalignment harmful action rate, lower is better (Loss of Control Evaluations table, section 4) | Alignment |
| 9 Jul 2026 | Muse Spark 1.1 system card | 1.1 | Agentic Misalignment (Lynch et al.) harmful action rate (Table 2, Table 13) | Alignment |
| 12 Jun 2026 | Muse Spark system card | 47.7 | Agentic Misalignment harmful action rate (Table 2, Table 23) | Alignment |
| 5 Feb 2026 | Claude Opus 4.6 system card | Agentic misalignment evaluations Research Sabotage, Framing for Crimes and Blackmail, misalignment rates under evaluation awareness steering and controls, text says near zero in all conditions, chart without printed values (6.5.4, Figure 6.5.4.A) | Alignment | |
| 24 Nov 2025 | Claude Opus 4.5 system card | Agentic misalignment evaluations Research Sabotage, Framing for Crimes and Blackmail, control and steered misalignment rates in an evaluation awareness steering case study, chart without printed values (Figure 6.7.2.B) | Alignment | |
| 29 Sep 2025 | Claude Sonnet 4.5 system card | 0 | Agentic misalignment suites, about 52,000 prompts, half with extended thinking. No deceptive actions with extended thinking, 2 without. Blackmail Suite, Research Sabotage and Framing for Crimes each 0.0 (7.5.4.1, Figure 7.5.4.1.A) | Alignment |
| 5 Aug 2025 | Claude Opus 4.1 system card addendum | Agentic misalignment blackmail scenario, extended thinking off and 16k, chart without printed values, in-training model (Figure 4.2.A) | Alignment | |
| 22 May 2025 | Claude Opus 4 system card | 84 | Opportunistic blackmail scenario, blackmail rate when the replacement AI shares its values: 84% of rollouts, lower is better (4.1.1.2) | Alignment |
More 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%
- StrongREJECT v2 (Meta) In use Models face a cascade of jailbreak attacks on disallowed behaviors adapted from StrongREJECT plus internal ones. Reports attack success rate. Muse Spark 1.1 0.50%