AI Agents Fail in Novel Ways, Put Businesses at Risk

Microsoft's AI Red Team identifies ten novel failure modes for agentic AI systems—such as agent compromise, indirect prompt injection, human-in-the-loop bypass, and memory poisoning—that can allow attackers to hijack agents or induce harmful actions; a memory-poisoning example shows how attacker-crafted emails can alter an agent's memory to exfiltrate data. The report urges defense-in-depth through threat modeling, clear trust boundaries, extensive logging and monitoring, automated assessments, and frequent red-teaming (and notes tools like PyRIT) to detect and mitigate these risks before and during production.