The security holes in AI agents that act on their own are now mapped

A new paper systematically maps the security weaknesses in AI agents that act on their own. This gives developers a clear list of what they need to fix to make these systems trustworthy.

Companies are already deploying AI agents that make their own plans and take actions. But the way these systems fail is fundamentally different from traditional software, and nobody had a clear way to secure them. This paper gives developers a first roadmap to build agent systems that can actually be governed, audited, and made resilient.