Computer use prompt injection evaluation (Anthropic)
Agents face about 600 prompt injection scenarios in coding platforms, browsers and email workflows. Score is the attack prevention rate.
In-house Anthropic evaluation reported with and without safeguards. The scenarios are not public.
- Made by
- Anthropic
- Kind
- Jailbreaks and prompt injection
- Level
- Expert
- Scored
- Graded by a model
- Answers
- Agent task
- Tasks
- Kept private
- Strongest model tested
- Claude Sonnet 4.5 #80 on the capability index
- Last in a system card
- October 2025
- Last model added
- October 2025
Claude Sonnet 4.5 leads with 78% among models released by October 2025
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.
The newest model with a score here came out in October 2025, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- Claude Fable 5.2
-
Gemini 4 Carbon - GPT-6 Astra
- Claude Opus 5.5
- and 16 more
- 1 Claude Sonnet 4.5 78%
- 2 Claude 3.7 Sonnet 74%
- 3 Claude Haiku 4.5 72.2%
- 4 Claude Opus 4 71%
- 5 Claude Sonnet 4 69%
Source: Labs’ model cards
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 78% | Sep 2025 | Lab model card | |
| 2 | Claude 3.7 Sonnet | 74% | Feb 2025 | Lab model card | |
| 3 | Claude Haiku 4.5 | 72.2% | Oct 2025 | Lab model card | |
| 4 | Claude Opus 4 | 71% | May 2025 | Lab model card | |
| 5 | Claude Sonnet 4 | 69% | May 2025 | Lab model card |
2 models have held number one in labs’ own numbers since 2023
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Labs’ model cards
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Labs’ model cards
Labs report it only in system cards
We read the headline results of 522 launch posts from 52 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 |
|---|---|---|---|---|
| Oct 15, 2025 | Claude Haiku 4.5 system card | 72.2 | Computer use prompt injection evaluation, attack prevention score without safeguards; 92.4 with safeguards (Table 3.2.2.A) | Robustness |
| Sep 29, 2025 | Claude Sonnet 4.5 system card | 78 | Computer use prompt injection evaluation, 70 to 110 hand-built cases, attack prevention score without safeguards. 82.6% with safeguards (Table 4.2.3.A) | Robustness |
| Aug 5, 2025 | Claude Opus 4.1 system card addendum | Prompt injection attacks and computer use, no number printed. Text says susceptibility very similar to Claude Opus 4 (3.2) | Robustness | |
| May 22, 2025 | Claude Sonnet 4 system card | 69 | Computer use prompt injection evaluation (about 600 scenarios), attack prevention score without safeguards 69%. With safeguards 86% (Table 3.2.A) | Robustness |
| May 22, 2025 | Claude Opus 4 system card | 71 | Computer use prompt injection evaluation (about 600 scenarios), attack prevention score without safeguards 71%. With safeguards 89% (Table 3.2.A) | Robustness |
| Feb 24, 2025 | Claude 3.7 Sonnet system card | 74 | Computer use prompt injection evaluation (176 tasks), prompt injections prevented with no mitigations 74%. With training and classifier mitigations 88% at a 0.5% false positive rate (4.2) | 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% - 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. Gemini 4 Carbon 0%
- DistressBench In use Models reply as a chatbot to 718 simulated users in psychological crisis, in single-turn and multi-turn chats. Three judge models score replies against clinician-written rubrics.
Muse Spark 1.2 88.4% - Lakera B3 (Backbone Breaker Benchmark) In use Crowdsourced prompt-injection attacks hit backbone LLMs inside agent threat snapshots at three defense levels. The score is attack resistance, 100 minus the board's vulnerability percentage. Claude Opus 4.5 89%
- IHEval In use Models face conflicts between system, user and tool instructions. Reports a composite score for following the higher-priority instruction. GPT-4o 70%