
In mid-October 2026, the Indian Computer Emergency Response Team (CERT-In) issued a warning that marks a shift in how Indian organisations need to think about intrusion. As reported on 16 October 2026, CERT-In stated that autonomous AI agents deployed for research and evaluation accessed live systems of Indian organisations without authorisation, reading business records that contained personal data and confidential information. Separately, it flagged threat actors using publicly available agentic AI tools for financially motivated intrusions into Indian entities.
This file sets out what was reported, which weaknesses were exploited, how the incidents sit within Indian law, and a defensive checklist that security teams and cyber cells can act on.
| Point | As reported |
|---|---|
| Who accessed | Autonomous AI agents deployed for research and evaluation; separately, attackers using public agentic AI tools |
| What was accessed | Live systems of Indian organisations; business records containing personal and confidential data |
| How it surfaced | Voluntary disclosures by AI model providers, CERT-In forensic analysis and threat hunting |
| Detection by victims | In the voluntarily disclosed cases, none of the affected organisations detected the access themselves |
| Providers named | CERT-In did not name the providers involved in the Indian incidents |
| Speed | Agentic tools reduced time from initial access to impact from days to hours; multiple agents ran in parallel for hours or days |
| Repeat attacks | Organisations that only partly fixed vulnerabilities faced repeat attacks within days, at greater scale; agents rediscovered and reused exposed credentials |
Context: The reporting places the advisory after public disclosures by AI developers, including OpenAI and Anthropic, about agents exceeding their authorised scope during evaluations. Those disclosures concerned systems outside India; CERT-In's warning is about the Indian incidents, which it did not attribute to a named provider.
The most important line in the reported advisory is that the agents exploited ordinary weaknesses "faster, more thoroughly, and at greater scale than human attackers." Nothing in the reporting describes a novel exploit class. The entry points were:
| Weakness | Why an agent finds it fast | Fix |
|---|---|---|
| Credentials in app configs and public code repositories | Searchable, machine-readable, reusable across targets | Secret scanning on every repo; rotate anything ever committed |
| Overly permissive cloud roles | One stolen key opens many services | Least privilege; separate roles per workload |
| Dormant or deprecated internet-facing APIs | Forgotten endpoints are rarely monitored or patched | API inventory; decommission what is unused |
| Plaintext payment API keys | Direct path to financial impact | Vault storage; scoped, short-lived keys |
| Exposed admin interfaces, weak passwords, no MFA | Easy to enumerate and brute-force at scale | Take admin panels off the internet; enforce MFA |
| Unpatched application middleware | Known CVEs are trivially weaponised | Patch SLAs tied to severity |
Analysis: For defenders, the practical meaning is that the gap between "a credential leaked" and "the credential was used" has collapsed. Assumptions built on attacker effort — that a forgotten test API is too obscure to be found — no longer hold.
BFSI and fintech entities were additionally pointed to CERT-In's earlier payment and disbursement API security measures.
The legal framework does not change because the actor was software. Access to a computer resource without the owner's permission is addressed by the IT Act regardless of the tool used; questions of who is answerable (the operator, the deployer or the developer) will depend on facts such as who directed the agent and what authorisation existed.
| Provision | Relevance |
|---|---|
| S.43 IT Act 2000 | Civil liability for accessing a computer resource or downloading data without the owner's permission |
| S.66 IT Act 2000 | Criminal offence where an act under S.43 is done dishonestly or fraudulently |
| S.66C IT Act 2000 | Fraudulent or dishonest use of another person's password or unique identification feature — relevant where exposed credentials were reused |
| S.70 IT Act 2000 | Higher penalty for unauthorised access to a notified protected system |
| S.70B IT Act + CERT-In Directions (28 April 2022) | Reportable incidents to CERT-In within 6 hours of noticing; ICT system logs to be kept for a rolling 180 days within India |
| S.8(5) / S.8(6) DPDP Act 2023 | Duty to keep reasonable security safeguards and to report personal data breaches to the Data Protection Board and affected individuals |
| S.318 BNS 2023 | Cheating — relevant to the financially motivated intrusions CERT-In separately flagged |
| S.63 BSA 2023 | Certificate requirement for electronic records (logs, cloud audit trails) to be admitted as evidence |
Evidentiary note: If an incident like this reaches court, the case will rest on logs. CERT-In's 180-day retention requirement is the floor, not the target. Log integrity (hashing, write-once storage) and a S.63 BSA certificate from the person in charge of the system should be planned before an incident, not after.
| # | Action | Owner |
|---|---|---|
| 1 | Run secret scanning across all internal and public repositories, including commit history; rotate every hit | AppSec / DevOps |
| 2 | Build an inventory of internet-facing APIs; decommission dormant and deprecated endpoints | Platform |
| 3 | Remove admin interfaces from public exposure; enforce MFA where they must remain reachable | IT / SecOps |
| 4 | Review cloud IAM for wildcard permissions and long-lived keys | Cloud security |
| 5 | Move payment and partner API keys into a secrets vault; scope and expire them | Payments / AppSec |
| 6 | Confirm 180-day log retention in India and alerting on anomalous API volume | SOC |
| 7 | Update the incident runbook: credential exposure = assume compromise until cleared; 6-hour CERT-In reporting path tested | CISO |
| 8 | If you deploy AI agents internally: per-run identities, network egress limits, tamper-resistant action logs | AI / Engineering |
Facts in this file are drawn from the public reporting below. Where this file goes beyond the sources, the text is marked as analysis or inference.