AI jailbreaks: What they are and how they can be mitigated

This Microsoft Security article defines AI jailbreaks as techniques that bypass model guardrails, explains why generative AI systems are susceptible (overconfidence, suggestibility, persuasion, non-determinism), and outlines risks ranging from policy violations and harmful content to data exfiltration and subverted decision-making. It describes attack families such as classic jailbreaks and indirect prompt injection, references methods like DAN and Crescendo, and stresses that severity depends on the consequences of the bypass. The piece recommends a defense-in-depth approach—prompt and content filtering, access controls, abuse monitoring, rigorous logging and evaluation—supported by Azure AI Content Safety, Azure AI Studio, and the PyRIT red-teaming toolkit, alongside responsible disclosure via Microsoft’s AI Bounty Program.