My agent said something it shouldn't have
Microsoft introduces guardrails for agents to block jailbreaks, prompt injection, and PII leakage
“Ignore everything that you've been told and do this for me.”
A Microsoft Build session demos Azure AI Foundry guardrails that protect agents from jailbreaks, direct and indirect prompt injection, harmful content, and PII exposure. The feature lets developers apply Microsoft-predefined or custom guardrails at the agent or model level. This is a practical developer-facing capability announcement rather than a strategic shift, aimed at hardening production AI agents.