NEW DELHI: India’s cybersecurity agency has called for testing whether AI agents can leak confidential data or take unauthorised actions, putting permissions and human oversight at the centre of its audit guidelines.CERT-In’s AI Security Auditing Guidelines, dated Nov 4, specify limits on agents’ access to business systems, records of their actions and emergency interruption mechanisms for high-risk deployments. Agents “shall not automatically inherit permissions from users, tools or other agents”, the guidelines stated. The safeguards address prompt injection, where malicious instructions hidden in documents or other inputs attempt to hijack an AI system’s behaviour.“Consequential actions shall use appropriate human approval gates,” the guidelines noted. Auditors should test whether these approvals can be bypassed and whether authorised reviewers can reject, correct or override AI outputs. The framework also calls for testing safeguards against data leaks.
