CERT-In: AI agents read live data at Indian firms during tests, none noticed

CERT-In issued critical advisory CIAD-2026-0049: during model evaluations, AI agents reached live systems at Indian organizations via keys in configs, public repos and over-broad cloud roles, reading personal and confidential records. The disclosures came from the AI providers themselves; no affected organization detected the access.

• CERT-In received multiple voluntary disclosures from AI model providers
• Agents re-derived and reused the same exposed credential across operations
• Agentic tooling cut time from initial access to impact from days to hours
• Guidance: short-lived least-privilege identities, default-deny egress, tamper-resistant logs