Have a safe flight! Building security guardrails for Copilot AI assistant

Natalia Semenova, a security architect at EPAM with prior experience at Microsoft and Google, draws on hands-on Microsoft Copilot deployments to walk through where guardrails belong in the Copilot shared-responsibility model. Since Copilot runs as a set of service accounts, access logs can obscure which end user actually touched sensitive data unless prompt-and-response logging is explicitly enabled. She maps controllable choke points onto the architecture: input filtering to block prohibited queries, prompt-injection detection, and output filtering or data loss prevention to catch leaked salaries, addresses or other labeled data before it reaches the user. Data labeling is the real precondition for safety, since an unlabeled file gets surfaced to anyone regardless of access rights, including through Copilot's own file-sharing suggestions. She covers consent and right-to-be-forgotten problems for PII embedded in a vector index, recommends anonymizing and separately encrypting data for retrieval-augmented generation, and points to native Azure reports and frameworks like MITRE ATLAS mapped to NIST as audit substitutes while formal AI compliance certifications remain unavailable. A long Q&A covers custom DLP implementation, multi-cloud non-human identity risk from a real client incident, stale documentation causing hallucinations, training-data guarantees, Copilot Web versus enterprise licensing, and zero trust login practices.