Earlier this month, Anthropic released Claude, an AI agent built to carry out tasks on a user’s computer. This innovative tool allows users to delegate various actions within their web browsers, significantly enhancing productivity and efficiency. By embodying the utility of an interactive assistant, Claude marks a major leap forward for AI applications.
Claude also has the capability to perform actions in a web browser on behalf of its users. This feature automates steps that often involve human touchpoints. This can mean navigating virtual learning, completing online assignments, or even just accessing email. Before doing any dangerous operations, Claude will first request your permission. This ranges from posting on social media to buying things to sharing sensitive information. This new safety feature is intended to reduce possible harms that could come from automated actions.
Anthropic rolled out Claude’s browsing ability to a limited number of 1,000 Max plan subscribers. These subscribers, who pay on average between $100 and $200 per month. This limited release gives Anthropic the opportunity to get immediate feedback and iterate on Claude’s capabilities and functionalities through real world interaction. Users have the option to restrict Claude’s access to specific websites through the app’s settings, enhancing control over the AI’s browsing activities.
To its credit, Anthropic has moved to strengthen their safety and security. Furthermore, they have put in place safeguards that block Claude from reading certain types of websites by default. These bans extend to platforms that discuss financial services, pornography and pirated content. Such measures reflect the company’s awareness of the evolving landscape of AI agents with browser capabilities and the inherent risks they may pose.
Even with these precautions taken, the emergence of AI agents into web spaces has already prompted conversations about these agents’ vulnerabilities. Last month, Brave’s security team stumbled upon an exciting discovery. Specifically, they discovered that Comet, an AI agent created by Perplexity, was susceptible to indirect prompt-injection attacks. In reaction to such studies, Anthropic has proposed a number of defenses against these types of attacks. Their interventions have remarkably reduced the success rate of prompt injection attempts. Secondly, it fell from 23.6% down to a paltry 11.2%, something indicating their deep prioritization of user safety.
Indeed, the landscape for competitive AI-powered browsers is becoming quite the hotbed of activity. Just last month, Perplexity hit the news when they announced an unsolicited $34.5 billion takeover bid for Chrome. This first electrifying move is illustrative of the greater tech sector’s swelling ambitions. Sam Altman, OpenAI’s CEO, is reportedly very keen on buying Chrome. The shift underscores the growing strategic importance of browser technology within that larger generative AI landscape.
At the same time, we hear that OpenAI is about to announce its own, similar, AI-enabled browser. Simultaneously, Perplexity has already launched Comet, which includes an AI helper that can actually help users with their tasks in a variety of ways. These advancements are encouraging signs that the market for AI-driven solutions is moving incredibly fast.
