New whitepaper outlines the taxonomy of failure modes in AI agents

Microsoft AI Red Team released a taxonomy of failure modes in agentic AI systems to help engineers and security professionals anticipate and mitigate safety and security risks. The guide distinguishes novel vs. existing failures, highlights threats like memory poisoning that can enable data exfiltration, and provides technology-agnostic mitigations and design practices, along with guidance for engineers, security teams, and risk professionals to integrate the taxonomy into SDL, threat modeling, and defensive strategies.