IT breach and the missing guardrails

AI companies have come forward with an awkward admission: Programs they were testing had strayed beyond the tasks allotted and found their way into the systems of Indian organisations

Somewhere in India, some machine may have read a file. It is possible your name and other details were on it. Clearly, you don’t know about it. Worse still, the company that holds the file may not know either. We know this happened because on October 15, CERT-In, the government’s cyber-emergency team, issued a warning it rated as critical.

IT breach and the missing guardrails.

AI companies have come forward with an awkward admission: Programs they were testing had strayed beyond the tasks allotted and found their way into the systems of Indian organisations. There, they read business records, including personal data.

You may have heard of these as AI agents: software that does not just answer questions, but can plan, run commands, research, and do everything else it is prompted to do. Ironically, these Agents did not break-in, rather, organisations had left the equivalent of ‘staff badges’ lying around: passwords in publicly viewable code, old test systems nobody had switched off, online storage that let almost anyone read almost anything. CERT-In calls these weaknesses conventional.

But the blunt word for ‘conventional’ is ‘outdated’. The AI Agents found those badges and got into the systems of organisations like any other human attacker would—but more efficiently because AI is faster.

So why did nobody notice the breach? Jaydeep Singh, General Manager (India Region) of the cyber security firm Kaspersky says an agent usually works through legitimate tools and valid credentials and its activity “can closely resemble that of a regular user.” It walks in wearing a staff badge.

This is why, Sunil Sharma of Sophos, another cybersecurity firm, says 81 per cent of ransomware attacks in India, the kind that lock a company’s files for money, begin with a hijacked login.

Now, here is the twist. Another advisory issued earlier has it that Indian companies must report a cyber incident to CERT-IN within six hours of noticing it, or of being told about it. Unauthorised access to computer systems is on the list. If they don’t, they can be penalised. But what does a company do when it does not know that there is a rogue agent in its system?

There are other questions too which add to the ambiguities. Does the rule bind AI companies whose programs (or agents) got in? The advisory covers any service provider, intermediary or “body corporate” that notices an incident. Whether that covers AI labs which may be based abroad is unclear. The advisory itself asks labs to tell the affected organisation and CERT-In to act promptly. But there is a difference between an advisory and a law.

Then comes the question that matters to the lay user: what happens to our data that gets compromised? The labs found the intrusions, so they know where their programs went, and they reported it in good faith. But good faith is not a law either.

India’s new privacy laws should help when implemented. Once it takes effect, a company that loses personal data must tell each affected person without delay, and send the new Data Protection Board a detailed report within 72 hours. But those duties start only in May 2027.

Meanwhile, the advisory issued this month carries another warning. Criminals are using free, openly available AI tools to break into Indian companies for money. What once took days, from first entry to final damage, now takes a few hours. And firms that patched only part of the hole were hit again within days.

As things are, this event has been filed as an AI problem. But this is actually a ‘housekeeping’ problem. The badges were left out of the room. No one was watching. The system is hopelessly careless. The only reason we know someone got in is because the people who knew about the unwanted visitors told the security agencies about it. So, if a file with your name on it was read, you are relying on somebody choosing to tell you. That should be unacceptable to us.