Updating the taxonomy of failure modes in agentic AI systems: What a year of red teaming taught us

After 12 months of red teaming, a Microsoft Security update reports a surge in real-world attacks on agentic AI systems and adds seven new failure modes, including supply chain compromise and goal hijacking. It outlines practical mitigations teams should adopt now.