Prompt injection robustness against adaptive attackers, browser use (Anthropic)
Claude Sonnet 5.5 0% leads since Sep 2026 Only in system cards
Browser agents face indirect prompt injections in web environments from an internal adaptive attacker. Score is attack success rate.
In-house Anthropic evaluation. The tasks are not public.
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
2 models share the top score of 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 Sonnet 5.5 0%
- 2 Claude Haiku 5.5 0%
- 3 Claude Opus 5.5 0.09%
- 4 Claude Sonnet 4.6 0.24%
- 5 Claude Opus 4.6 0.29%
- 6 Claude Mythos Preview 0.68%
- 7 Claude Opus 4.7 0.74%
- 8 Claude Sonnet 5 0.93%
- 9 Claude Fable 5.1 2.64%
- 10 Claude Opus 5 3.84%
- 11 Claude Opus 4.5 8.00%
- 12 Claude Fable 5 29.7%
- 13 Claude Opus 4.8 31.5%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Sonnet 5.5 | 0% | Sep 2026 | Model card | |
| 2 | Claude Haiku 5.5 | 0% | Oct 2026 | Model card | |
| 3 | Claude Opus 5.5 | 0.09% | Sep 2026 | Model card | |
| 4 | Claude Sonnet 4.6 | 0.24% | Feb 2026 | Model card | |
| 5 | Claude Opus 4.6 | 0.29% | Feb 2026 | Model card | |
| 6 | Claude Mythos Preview | 0.68% | Apr 2026 | Model card | |
| 7 | Claude Opus 4.7 | 0.74% | Apr 2026 | Model card | |
| 8 | Claude Sonnet 5 | 0.93% | Jun 2026 | Model card | |
| 9 | Claude Fable 5.1 | 2.64% | Sep 2026 | Model card | |
| 10 | Claude Opus 5 | 3.84% | Jul 2026 | Model card | |
| 11 | Claude Opus 4.5 | 8.00% | Nov 2025 | Model card | |
| 12 | Claude Fable 5 | 29.7% | Jun 2026 | Model card | |
| 13 | Claude Opus 4.8 | 31.5% | May 2026 | Model card |
5 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.
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 | Browser use in Claude Cowork with auto mode, attempt-level attack success rate, 0/110 scenarios (Table 5.2.2.3.A) | Robustness |
| 28 Sep 2026 | Claude Sonnet 5.5 system card | 0 | Browser use prompt injection in Claude Cowork, attempt-level attack success rate without safeguards, 0/110 scenarios (0% with auto mode) (Table 5.2.2.3.A) | Robustness |
| 22 Sep 2026 | Claude Opus 5.5 system card | 0.09 | Browser use prompt injection in the Claude Cowork harness, attempt-level attack success rate without safeguards, with thinking at high effort, 1/110 scenarios (0% with auto mode), lower is better (Table 5.2.2.3.A) | Robustness |
| 1 Sep 2026 | Claude Fable 5.1 system card | 2.64 | Browser use prompt injection in the Claude Cowork harness, attempt-level attack success rate, thinking at high effort, without PI probes, with fallbacks; 10/110 scenarios; 0% with auto mode (Table 5.2.2.3.A) | Robustness |
| 24 Jul 2026 | Claude Opus 5 system card | 3.84 | Browser use prompt injection in the Claude Cowork harness, attempt-level attack success rate without safeguards, thinking at medium effort, 11/129 scenarios (0% with auto mode), lower is better. The prose says 3.70%. Updated harness from the August 19 revision (5.2.2.4, second Table 5.2.2.3.A) | Robustness |
| 30 Jun 2026 | Claude Sonnet 5 system card | 0.93 | Browser use prompt injection, 129 environments, attempt-level attack success rate without safeguards, with thinking, 9/129 scenarios (1.01% without thinking. 0% with safeguards), lower is better (Table 5.2.2.4.A) | Robustness |
| 9 Jun 2026 | Claude Fable 5 system card | 29.7 | Professional red-teamer prompt injection in browser use, 129 environments x10 attempts, attempt-level attack success rate without safeguards, with thinking, row labeled Mythos 5, 71/129 scenarios; deployed safeguards 6.5% (25/129), updated safeguards 0% (0/129), lower is better (Table 5.2.2.3.A) | Robustness |
| 28 May 2026 | Claude Opus 4.8 system card | 31.5 | Browser use prompt injection from professional red teamers, 129 held-out environments, per-attempt attack success rate without safeguards, with thinking (62.8% of scenarios. With safeguards 0.5% of attempts and 3.9% of scenarios), lower is better (Table 5.2.2.4.A) | Robustness |
| 16 Apr 2026 | Claude Opus 4.7 system card | 0.74 | Browser use prompt injection from professional red-teamers, with thinking, without safeguards, per-attempt ASR over 148 environments. 4.05 of scenarios. 0.00 with safeguards (Table 5.2.2.3.A) | Robustness |
| 7 Apr 2026 | Claude Mythos Preview system card | 0.68 | Browser use, professional red-teamer attacks sourced against Opus 4.6, share of 148 environments with at least 1 successful attack (1/148), extended thinking, without safeguards, lower is better; 0.00 with safeguards (Table 8.3.2.2.3.A) | Robustness |
| 17 Feb 2026 | Claude Sonnet 4.6 system card | 0.24 | Best-of-N browser prompt injection, extended thinking, without safeguards, per-attempt ASR; 1.29 of scenarios, 0.08 per attempt with updated safeguards (Tables 5.2.2.3.A and 5.2.2.3.B) | Robustness |
| 5 Feb 2026 | Claude Opus 4.6 system card | 0.29 | Best-of-N browser prompt injection, 389 scenarios, extended thinking, without safeguards, per-attempt attack success rate, lower is better; 2.06 of scenarios (Table 5.2.2.3.A); standard thinking with updated safeguards 0.08 per attempt and 0.77 of scenarios (Table 5.2.2.3.B) | Robustness |
| 24 Nov 2025 | Claude Opus 4.5 system card | 8 | Claude for Chrome internal adaptive evaluation, attack success rate, model only without safeguards, 100 attempts per environment, lower is better (Figure 5.2.2.3.A; 7.6% with old safeguards, 1.4% with improved 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%