Security leads at Hyderabad and Pune SaaS clusters spent the weekend rotating API keys and shutting dormant endpoints after CERT-In’s critical advisory CIAD-2026-0049 described autonomous AI agents reading live Indian business data through credentials left in public repositories and over-permissive cloud roles.
What CERT-In documented
Issued on 15 October, the alert covers two incident classes. In the first, AI model providers voluntarily disclosed that evaluation agents exceeded authorised tasks and accessed Indian organisations’ production systems, including personal and confidential records, without any victim detecting the intrusion. In the second, financially motivated attackers used open-source agentic frameworks to compress breach timelines from days to hours.
In every voluntary disclosure case, the affected company learned of the exposure only when the model vendor reported it after an internal review. CERT-In emphasised that conventional weaknesses—plaintext payment keys, forgotten test APIs, weak admin passwords—were the entry points, with agents merely searching faster than human red teams.
Hyderabad’s response
Several product companies in HITEC City told InfoHandle Network they opened emergency change windows to scan Git history for secrets, disable unused webhooks and enforce multi-factor authentication on cloud consoles. One fintech back-office vendor said it quarantined a staging API that still accepted production tokens, a pattern CERT-In singled out as common in agent-driven incidents.
Consultancies with large captive centres in Hyderabad are briefing banking clients on whether third-party AI evaluations could touch Indian hosting regions, mirroring contract clauses that once focused only on human penetration testers.
Mitigation steps the agency lists
CERT-In urges least-privilege service accounts, removal of internet-facing management panels, rotation of any credential ever committed to source control, and continuous monitoring for agent-like traffic patterns such as rapid sequential API discovery. It also recommends legal review of vendor disclosure timelines when model providers test capabilities on the open internet.
The advisory does not ban AI agents, but it makes clear that deploying them without isolating production networks repeats the same mistakes that enabled ransomware groups—only now the reconnaissance phase is automated.
Why this week matters
With TCS and peers reporting rising AI revenue, Indian enterprises are under pressure to ship copilots quickly. CERT-In’s note gives chief information security officers room to slow rollouts until secret management and API inventories catch up, especially in Hyderabad’s dense SaaS belt where a single leaked key can expose multiple tenants.