1/20/26 A prompt-injection vulnerability disclosed recently in Anthropic’s Cowork AI highlights ongoing risks in agentic AI tools. Security researchers showed how a hidden prompt can trick Cowork into exfiltrating sensitive files to an attacker’s account once file access is granted.
As CISOs enable non-technical office workers to use AI productivity tools such as Cowork, they should be aware that Anthropic’s mitigation advice places the burden on users to recognize and prevent suspicious actions.