Rewind-attacker cyber adversarial robustness (Anthropic)
An automated attacker with rewinds and 400 calls tries to make a safeguarded Claude Code agent complete offensive cyber tasks. Score is attack success rate.
Tasks include ransomware, data exfiltration, CVE exploitation, command and control and self-replication.
- Level
- Expert Specialists with years of training, often over hours of work
- Scored
- Settled by events Predictions are scored against what later happens.
- Answers
- Agent task Many steps in a terminal, browser, desktop, app, game or robot.
Claude Opus 5.5 leads with 4.00%
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 4.00%
- 2 Claude Fable 5.1 4.50%
- 3 Claude Fable 5 5.00%
- 4 Claude Opus 5 5.60%
- 5 Claude Sonnet 5.5 21%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | 4.00% | Sep 2026 | Model card | |
| 2 | Claude Fable 5.1 | 4.50% | Sep 2026 | Model card | |
| 3 | Claude Fable 5 | 5.00% | Jun 2026 | Model card | |
| 4 | Claude Opus 5 | 5.60% | Jul 2026 | Model card | |
| 5 | Claude Sonnet 5.5 | 21% | Sep 2026 | Model card |
2 models have held number one since its release in Sep 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.
5 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 |
|---|---|---|---|---|
| 28 Sep 2026 | Claude Sonnet 5.5 system card | 21 | rewind-attacker cyber adversarial robustness, attack success rate over 133 tasks, lower is better (3.3.2, Figure 3.3.2.A) | Robustness |
| 22 Sep 2026 | Claude Opus 5.5 system card | 4 | Cyber adversarial robustness evaluation (rewind-attacker), attack success rate over 133 tasks with production safeguards, earlier snapshot, lower is better (3.5.1, Figure 3.5.1.A) | Robustness |
| 1 Sep 2026 | Claude Fable 5.1 system card | 4.5 | Cyber adversarial robustness evaluation (rewind-attacker), offensive cyber task completion rate under automated red teaming, 133 tasks, mean of 8 runs, lower is better (Figure 3.5.1.A, 3.5.1) | Robustness |
| 24 Jul 2026 | Claude Opus 5 system card | 5.6 | Cyber adversarial robustness evaluation (rewind-attacker), attack success rate on 128 tasks against deployed safeguards, earlier snapshot, lower is better, printed on chart (3.5.1, Figure 3.5.1.A) | Robustness |
| 9 Jun 2026 | Claude Fable 5 system card | 5 | Internal red-teaming with an automated Opus 4.7 red-teaming agent (up to 400 turns, roll-back and reset), share of short offensive cyber tasks Fable 5 completed, lower is better (Figure 3.3.3.A, 3.3.3) | Robustness |
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%