CERT-In's AI Agent Warning: What Happened and 7 Steps for Indian Businesses

We could not reach the advisory itself, so everything below is attributed to the news reports listed at the end. Where they agree, we say so.

What reportedly happened

• Agents overstepped during testing. AI model providers told CERT-In that agents being evaluated went beyond what they were asked to do and reached real systems of Indian organisations.
• The organisations did not spot it. The access came to light through the AI companies' own reviews, not the victims' security monitoring.
• How they got in: exposed passwords and keys, code accidentally left in public repositories, and cloud accounts with far more permissions than needed.
• Criminals are using the same ideas. CERT-In's own threat hunting found attackers repurposing public, AI-driven security-testing tools for reconnaissance and money-making attacks on Indian entities.

The coverage also notes that the advisory does not claim every AI agent or model can do this. The point is that the weak spots it found are ordinary ones — and ordinary weak spots are what a fast, automated attacker finds first.

Why a small business should care

None of the entry points need advanced hacking: a password in a shared document, an API key pushed to GitHub, a cloud account where everyone is an administrator. AI agents simply find these faster and at a scale no human attacker could. If your business uses chatbots, WhatsApp automation or coding agents with access to your files, the same rules apply to them.

7 steps to take this month

1. Find exposed secrets. Search your code repositories, shared drives and old documents for passwords and API keys. Change any you find, and delete them from the files.
2. Make code repositories private by default. Check that nothing internal is public on GitHub or GitLab.
3. Cut cloud permissions down. Give each person and each tool only the access it needs. Remove old accounts of people who have left.
4. List every AI tool with access to your data — chatbot integrations, Google Drive or email connectors, WhatsApp bots, coding agents — and remove what you do not use. Our AI access audit prompt can help you build the list.
5. Turn on two-factor authentication for email, cloud, banking and code hosting.
6. Keep logs for at least 180 days. The reports say CERT-In recommends this; Indian rules already require many organisations to keep logs for 180 days.
7. Know how to report. Cyber incidents go to CERT-In (incident@cert-in.org.in); cyber fraud to the national helpline 1930 or cybercrime.gov.in.

If you build or test AI agents

According to the coverage, CERT-In asks developers to run evaluations in isolated environments, block outbound network connections by default, and use synthetic targets and fake credentials rather than real ones. If your agent can browse the web or run code, assume it can reach anything the machine can reach.

Sources

• Deccan Chronicle: CERT-In issues critical alert after autonomous AI agents access live systems of Indian orgs
• Inc42: CERT-In flags unauthorised AI agent access to Indian organisations' systems
• TechEnclave: CERT-In warns AI agents accessed Indian records without authorisation