ClaudeBleed: Critical Flaw Hijacks Claude's Browser Extension

Cyber Updates - Asher Tamam reports on a critical vulnerability, dubbed ‘ClaudeBleed’ by LayerX researchers, affecting the Claude AI browser extension. This flaw allows any arbitrary browser extension, even one with zero permissions, to completely hijack the Claude extension. Attackers can inject commands, compelling Claude to perform autonomous actions on a user’s behalf.

According to Cyber Updates - Asher Tamam, the attack model is both elegant and stealthy. Exploitation could lead to unauthorized actions ranging from exfiltrating confidential files from Google Drive and stealing code from GitHub to sending emails without user consent. All of these actions can occur completely under the user’s radar, bypassing explicit interaction.

Anthropic’s initial ‘half-fix’ introduced a user approval prompt for certain actions. However, researchers demonstrated this control is easily circumvented by simply toggling the extension into ‘autonomous mode’ or via visual manipulation of Claude’s interface. Cyber Updates - Asher Tamam concludes that in the race for AI companies to deliver automation and productivity, fundamental security boundaries are being left completely exposed.