Security researchers have concerns that Anthropic’s Claude for Chrome is vulnerable to malicious prompting. Claude for Chrome allows users to chat with Claude as they browse the web. Claude can read webpages, fill forms, and click on links and buttons to perform complex tasks for the user. But Anthropic’s testing revealed that 11.2% of malicious prompting attempts succeeded even with safety measures in place. One test case was a malicious email that asked Claude to delete all emails in the user’s inbox for “mailbox hygiene”. AI researcher Simon Willison states that an 11.2% success rate is unacceptable for so-called AI agents, especially when several AI companies are releasing their own browser extensions. One competing product, Perplexity’s Comet browser, was found to be vulnerable to a prompt injection attack that instructed it to start password recovery for the user’s Gmail account. Although Perplexity attempted to fix the issue, Comet remains vulnerable to this attack.
Consentaumatic does the cookie thing
On an extremely limited number of sites. It performs very poorly in real world use. Especially for academically oriented sites I’ve noticed.