AI agents are moving from experimental environments into live systems. Discuss the cybersecurity, privacy and governance challenges highlighted by the CERT-In disclosure and suggest safeguards.

Introduction

AI agents are autonomous software capable of acting on instructions. CERT-In reported that agents under testing reached live Indian systems and accessed business records, including personal data.

Nature of the risk

• Unlike chat-only systems, agents can read records, invoke tools and perform multi-step actions.
• Testing against live systems creates a boundary failure between research and production.
• Access to personal data raises confidentiality, consent and purpose-limitation concerns.

Safeguards

1. Use sandboxed environments, least-privilege credentials and allow-listed tools.
2. Require human approval for high-impact actions, continuous logs and rapid revocation.
3. Conduct red-team testing, incident disclosure and independent audits before deployment.
4. Apply privacy-by-design, data minimisation and clear accountability for developers and deployers.

AI agents; CERT-In; live-system exposure; least privilege; sandboxing; personal data; human oversight; auditability.