Prompt injection robustness against adaptive attackers, computer use (Anthropic)
Claude Opus 4.5 0% leads since Nov 2025 Only in system cards
Computer-use agents face indirect prompt injections crafted by an adaptive attacker. Score is attack success rate.
Anthropic uses Gray Swan's Shade adaptive attacker for this surface. Reported with and without safeguards.
Made by Anthropic Released Nov 2025 Jailbreaks and prompt injection 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
- Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
- Answers
- Agent task Many steps in a terminal, browser, desktop, app, game or robot.
13
Models scored in lab cards
13
System cards report it, from 1 lab
Claude Opus 4.5 leads with 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 Opus 4.5 0%
- 2 Claude Fable 5.1 0.07%
- 3 Claude Opus 5.5 0.07%
- 4 Claude Sonnet 5.5 0.07%
- 5 Claude Haiku 5.5 0.07%
- 6 Claude Mythos Preview 0.43%
- 7 Claude Opus 4.7 0.46%
- 8 Claude Opus 5 0.54%
- 9 Claude Fable 5 0.82%
- 10 Claude Sonnet 5 2.25%
- 11 Claude Opus 4.8 7.14%
- 12 Claude Sonnet 4.6 12%
- 13 Claude Opus 4.6 17.8%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 4.5 | 0% | Nov 2025 | Model card | |
| 2 | Claude Fable 5.1 | 0.07% | Sep 2026 | Model card | |
| 3 | Claude Opus 5.5 | 0.07% | Sep 2026 | Model card | |
| 4 | Claude Sonnet 5.5 | 0.07% | Sep 2026 | Model card | |
| 5 | Claude Haiku 5.5 | 0.07% | Oct 2026 | Model card | |
| 6 | Claude Mythos Preview | 0.43% | Apr 2026 | Model card | |
| 7 | Claude Opus 4.7 | 0.46% | Apr 2026 | Model card | |
| 8 | Claude Opus 5 | 0.54% | Jul 2026 | Model card | |
| 9 | Claude Fable 5 | 0.82% | Jun 2026 | Model card | |
| 10 | Claude Sonnet 5 | 2.25% | Jun 2026 | Model card | |
| 11 | Claude Opus 4.8 | 7.14% | May 2026 | Model card | |
| 12 | Claude Sonnet 4.6 | 12% | Feb 2026 | Model card | |
| 13 | Claude Opus 4.6 | 17.8% | Feb 2026 | Model card |
Number one over time
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
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.
13 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 | 0.07 | Shade adaptive attacker, computer use GUI, attempt-level attack success rate without safeguards, 1/14 scenarios. Same with probes (Table 5.2.2.2.A) | Robustness |
| 28 Sep 2026 | Claude Sonnet 5.5 system card | 0.07 | Shade adaptive attacker in computer use, attempt-level attack success rate without safeguards, with thinking, 1/14 scenarios (0.07% with probes) (Table 5.2.2.2.A) | Robustness |
| 22 Sep 2026 | Claude Opus 5.5 system card | 0.07 | Shade adaptive attacker in computer use, attempt-level attack success rate without safeguards, with thinking, 1/14 scenarios (0.04% with probes), lower is better (Table 5.2.2.2.A) | Robustness |
| 1 Sep 2026 | Claude Fable 5.1 system card | 0.07 | Shade indirect prompt injection in computer use, attempt-level attack success rate, with thinking, without PI probes, with fallbacks; 2/14 scenarios; 0.07% with PI probes (Table 5.2.2.2.A) | Robustness |
| 24 Jul 2026 | Claude Opus 5 system card | 0.54 | Shade adaptive attacker in computer use, attempt-level attack success rate without safeguards, with thinking, 1/14 scenarios (0.25% with probes), lower is better (Table 5.2.2.3.A) | Robustness |
| 30 Jun 2026 | Claude Sonnet 5 system card | 2.25 | Shade adaptive attacker in computer use, attempt-level attack success rate without safeguards, with thinking, 4/14 scenarios (6.04% without thinking. 1.46% with safeguards), lower is better (Table 5.2.2.3.A) | Robustness |
| 9 Jun 2026 | Claude Fable 5 system card | 0.82 | Shade indirect prompt injection in computer use, attempt-level attack success rate without safeguards, with thinking, 14 scenarios x200 attempts, row labeled Mythos 5, 4/14 scenarios broken; with safeguards 0.46% and 3/14, lower is better (Table 5.2.2.2.A) | Robustness |
| 28 May 2026 | Claude Opus 4.8 system card | 7.14 | Shade adaptive attacker in computer use, 14 test cases, single-attempt attack success rate without safeguards, with thinking (200 attempts 50.0%. With safeguards 5.11% and 57.1%), lower is better (Table 5.2.2.3.A) | Robustness |
| 16 Apr 2026 | Claude Opus 4.7 system card | 0.46 | Shade attacks in computer use, with thinking, without safeguards, 1 attempt. 200 attempts 7.14. Without thinking 0.39 and 21.43. With safeguards 0.61 and 14.29 (Table 5.2.2.2.A) | Robustness |
| 7 Apr 2026 | Claude Mythos Preview system card | 0.43 | Shade attacks in computer use, extended thinking, without safeguards, 1 attempt, lower is better; 21.43 at 200 attempts, 0.32 and 21.43 with safeguards (Table 8.3.2.2.2.A) | Robustness |
| 17 Feb 2026 | Claude Sonnet 4.6 system card | 12 | Shade attacks in computer use, extended thinking, without safeguards, 1 attempt; 42.9 at 200 attempts, 8.0 and 50.0 with safeguards (Table 5.2.2.2.A) | Robustness |
| 5 Feb 2026 | Claude Opus 4.6 system card | 17.8 | Shade adaptive prompt injection in computer use, stronger attacker, extended thinking, without safeguards, 1 attempt, lower is better; 78.6 at 200 attempts, 9.7 and 57.1 with safeguards (Table 5.2.2.2.A); original attacker without thinking 0.29 and 7.14 (5.2.2.2) | Robustness |
| 24 Nov 2025 | Claude Opus 4.5 system card | 0 | Shade adaptive indirect prompt injection, computer use, attack success rate at 1 attempt, extended thinking, without safeguards, lower is better (Table 5.2.2.2.A; 0.0% with 200 attempts, standard thinking 28.6% without and 14.3% with safeguards) | 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%