Anthropic’s agentic browser: Claude for Chrome
In late August 2025, Anthropic launched a research preview of Claude for Chrome, an AI browser agent powered by its Claude models. The company explained that giving Claude the ability to “see what you’re looking at, click buttons, and fill forms” in the browser would make the assistant far more useful, but it also acknowledged safety challenges. The pilot program started with 1,000 users on the high‑tier Max plan and a waitlist for others. Users can chat with Claude in a sidecar window that maintains context and, with permission, allows the agent to take actions on their behalf.
Safety is a central concern. Anthropic’s internal “red‑teaming” experiments showed a 23.6 % prompt‑injection success rate without mitigations, meaning malicious instructions hidden in emails or websites could trick the agent into harmful actions. By adding site‑level permissions, action confirmations and improved system prompts, Anthropic said it reduced the attack success rate to 11.2 %. The company has also blocked Claude from accessing high‑risk categories such as financial services and adult content and is building classifiers to detect hidden malicious instructions. TechCrunch notes that the launch comes amid a broader race to build AI‑powered browsers — competitors like Perplexity’s Comet and Google’s Gemini integrations are already vying for user attention.
