Who is responsible when AI agents go out of control?

AI-assisted source summary · not independently verified. Content standards

ClawSecure tested 14 models from five labs and found recurring security issues in the models and their integrations with applications such as Dropbox, Notion, and Linear. Attackers can hide instructions in emails, shared documents, or support tickets to induce agents with access to financial accounts to take action. ClawSecure’s founder argues that providers should explain the limits of their safeguards and establish public, comparable security ratings.