Anthropic is launching its browser-based AI assistant, Claude, in a controlled testing phase for Google Chrome users. The extension aims to let Claude work directly within the browser, providing a more seamless experience for users.
However, this development also brings safety and security challenges that need to be addressed. To overcome these concerns, Anthropic is implementing various safeguards, including site-level permissions, action confirmations, and additional protections against prompt injection attacks.
Prompt injection attacks, where malicious actors hide instructions in websites or emails to trick AIs into performing harmful actions without user knowledge, are a significant risk. Anthropic has conducted extensive testing, including red-teaming experiments, which found that browser use without safety mitigations resulted in a 23.6% attack success rate.
To combat this threat, Claude’s new defenses include improved system prompts, blocked access to high-risk websites, and advanced classifiers to detect suspicious instruction patterns. The company has also reduced the attack success rate from 23.6% to 11.2% with these safety measures in place.
Anthropic is now seeking trusted testers to participate in a research preview of Claude for Chrome. Users can join the waitlist at claude.ai/chrome and install the extension once access is granted. The company encourages users to start with trusted sites, avoid sensitive information, and provide feedback to help improve the capabilities and safeguards of Claude for Chrome.
By launching this browser-based AI assistant, Anthropic aims to integrate AI into our lives in a fundamentally new way while prioritizing safety and security.
Source: https://www.anthropic.com/news/claude-for-chrome