Training distribution hack rate (Anthropic)
Classifier monitors measure reward hacking on a subset of Anthropic's reinforcement learning training episodes.
Monitoring metric on training data, not a fixed public task set.
- 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.
Claude Opus 5.5 leads with 0.63%
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 5.5 0.63%
- 2 Claude Opus 4.5 1.00%
- 3 Claude Sonnet 4.5 2.00%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | 0.63% | Sep 2026 | Model card | |
| 2 | Claude Opus 4.5 | 1.00% | Nov 2025 | Model card | |
| 3 | Claude Sonnet 4.5 | 2.00% | Sep 2025 | Model card |
3 models have held number one since its release in Nov 2025
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 |
|---|---|---|---|---|
| 22 Sep 2026 | Claude Opus 5.5 system card | 0.63 | Successful reward hacks, share of RL training episodes flagged by a general reward-hack classifier on environments shared with Opus 5 and Mythos 5.1, lower is better (6.2.1, Figure 6.2.1.A) | Alignment |
| 24 Nov 2025 | Claude Opus 4.5 system card | 1 | Subset of training data, classifier hack rate, lower is better (Table 6.10.1.A) | Alignment |
| 29 Sep 2025 | Claude Sonnet 4.5 system card | 2 | Training distribution classifier hack rate, environ 1, 2%, lower is better (Table 6.1.A) | Alignment |
| 5 Aug 2025 | Claude Opus 4.1 system card addendum | Training distribution classifier hack rate, environ 1 10% and environ 2 3%, no overall (Table 5.B) | Alignment | |
| 22 May 2025 | Claude Sonnet 4 system card | Training distribution classifier hack rate, environ 1 13% and environ 2 2%, no overall (Table 6.2.A) | Alignment | |
| 22 May 2025 | Claude Opus 4 system card | Training distribution classifier hack rate, environ 1 15% and environ 2 2%, no overall (Table 6.2.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%