Anthropic undermining liberal democracy (behavioral audit metric)
Automated behavioral audit metric: investigator models probe willingness to help coups, election tampering, or political surveillance, censorship or intimidation.
One metric in Anthropic's automated behavioral audit (~2,900 investigations from ~1,450 seeds, two investigator models, 95% CIs). Card footnote says it was inspired by Free Systems' Dictatorship Eval. Scores appear only in charts.
- Scored
- Graded by a model A language model scores each response, usually against a rubric.
- Answers
- Open-ended Free-form text, documents, designs or media judged as a whole.
Claude Opus 5 leads with 1.13
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 Opus 5 1.13
- 2 Claude Opus 5.5 1.17
- 3 Claude Fable 5 1.20
- 4 Claude Sonnet 5.5 1.21
- 5 Claude Sonnet 5 1.22
- 6 Claude Haiku 5.5 1.22
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 5 | 1.13 | Jul 2026 | Model card | |
| 2 | Claude Opus 5.5 | 1.17 | Sep 2026 | Model card | |
| 3 | Claude Fable 5 | 1.20 | Jun 2026 | Model card | |
| 4 | Claude Sonnet 5.5 | 1.21 | Sep 2026 | Model card | |
| 5 | Claude Sonnet 5 | 1.22 | Jun 2026 | Model card | |
| 6 | Claude Haiku 5.5 | 1.22 | Oct 2026 | Model card |
2 models have held number one since its release in May 2026
Lab claims count from the day the card came out, independent runs from the model’s release.
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
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.
6 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.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| 7 Oct 2026 | Claude Haiku 5.5 system card | 1.22 | Automated behavioral audit, undermining liberal democracy score on a 1 to 10 scale, lower is better (Figure 6.2.1.A) | Alignment |
| 28 Sep 2026 | Claude Sonnet 5.5 system card | 1.21 | Automated behavioral audit, undermining liberal democracy score on a 1 to 10 scale, lower is better (6.2.1, Figure 6.2.1.B) | Alignment |
| 22 Sep 2026 | Claude Opus 5.5 system card | 1.17 | Automated behavioral audit, undermining liberal democracy score on a 1 to 10 scale, lower is better (6.4.1, Figure 6.4.1.A) | Alignment |
| 24 Jul 2026 | Claude Opus 5 system card | 1.13 | Automated behavioral audit, undermining liberal democracy score on a 1 to 10 scale, lower is better, printed on chart (6.4.1, Figure 6.4.1.A) | Alignment |
| 30 Jun 2026 | Claude Sonnet 5 system card | 1.22 | Automated behavioral audit, undermining liberal democracy score on a 1 to 10 scale, lower is better (6.4.1, Figure 6.4.1.A) | Alignment |
| 9 Jun 2026 | Claude Fable 5 system card | 1.2 | Undermining liberal democracy, safeguards-on automated behavioral audit with Fable 5 behind production classifiers and fallback, 1 to 10 scale, lower is better, printed on chart (Figure 6.2.3.2.A) | 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%
- 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%