According to @AnthropicAI, browser use introduces safety challenges for AI models—especially prompt injection—and the company has launched a pilot to strengthen existing defenses in Claude’s browsing capability (Source: @AnthropicAI, Aug 26, 2025). According to @AnthropicAI, the announcement shares that safety measures already exist and the pilot aims to improve them, while providing no timelines, metrics, product release details, or any cryptocurrency/market impact disclosures (Source: @AnthropicAI, Aug 26, 2025).

According to Berkeley AI Research, their latest blog post discusses innovative techniques to defend against prompt injection attacks using Structured Queries (StruQ) and Preference Optimization (SecAlign). These methods, led by Sizhe Chen and Julien Piet, aim to enhance AI model security by structuring queries to prevent unauthorized data manipulation and optimizing preferences to align with secure protocols.