PromptFiction Flaw Auto-Submitted Hidden Prompts in Claude Desktop

Oasis Security disclosed “PromptFiction,” a vulnerability in Anthropic’s Claude Desktop where the app’s claude:// URL handler accepted a q parameter and automatically submitted prompts when a user clicked crafted links. Attackers could hide malicious instructions in long prompts to exfiltrate previous conversations via the Files API, and—when filesystem integration was enabled—implant remote debugging code or persistence in user files, potentially enabling code execution. Anthropic fixed the behavior so prompts are pre-filled and require user review in Claude Desktop version 1.1.2321 or later.