Microsoft Security released updated research on failure modes in agentic AI systems based on twelve months of red-team testing, identifying seven new attack vectors including supply chain compromise and goal hijacking, along with mitigation recommendations.