Claude for Chrome (research preview)
Claude for Chrome starts as a research preview with 1,000 Max users, cutting prompt-injection attack success from 23.6% to 11.2% with mitigations.
A browser extension where Claude acts in the user's browser. Anthropic tested 123 attack cases across 29 scenarios, with 23.6% success without mitigations and 11.2% with them in autonomous mode. On four browser-specific attacks the rate fell from 35.7% to 0%. Added site permissions, action confirmations, and blocks on high-risk site categories.
- Date
- Monday, 25 August 2025
- Lab
- Anthropic
- Kind
- product
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Prompt-injection attack success rate | 23.6% -> 11.2% with safety mitigations, autonomous mode; 123 test cases | company |
| Browser-specific attack success rate | 35.7% -> 0% four-attack challenge set | company |
Expanded to all Max users 2025-11-24 and to Pro, Team and Enterprise 2025-12-18 (page updates). Attack-rate figures are Anthropic-run.
Sources
This record was checked against its sources on 6 October 2026. How we check