- Published on
- Published
A recently discovered vulnerability in Anthropic's Claude Cowork allowed local AI agents to escape their virtual playpens and wander unsupervised through their hosts' private files, proving that even the most polite algorithms cannot resist a good snoop.