Rogue AI agents are here, experts say Indian companies can no longer assume internal systems are safe

Rogue AI agents have officially entered India. Through meticulous planning and execution, they have managed to breach the seemingly secure walls of Indian companies accessing their confidential data, CERT-In has informed. It is a first for India. Incidents of rogue AI activity are being reported worldwide.

In Short

• CERT-In warns AI agents accessed Indian organisations
• Experts say AI agents spreading like virus on the internet
• Indian companies need stricter measures to tackle this threat

India’s cybersecurity agency, Computer Emergency Response Team (CERT-In), recently issued an advisory. Rated, “critical,” the advisory said that several Indian organisations had had their confidential data exposed to rogue AI agents.

These AI agents, CERT-In says, managed to gain access to the systems through vulnerabilities that exposed credentials and other settings of the companies. The more alarming part of this incident is that the affected Indian organisations – that have not been named yet – did not identify these breaches. “The incidents came to light through voluntary disclosures by AI model providers,” the CERT-in mentions. “The affected organisations could not detect the unauthorised access on their own.”

It is likely that you recognise that this could be a major risk for data as these agents were able to access “business records, including personal data and confidential material.” However, the story goes a bit deeper. To explain that to you, we may have to take a trip back in time.

A blast from the past

Back in 1997, when the internet was still a new thing – and not as embedded into our lives – it faced a problem called viruses. Computer viruses seemed to be everywhere. "The only truly secure system is one that is powered off, cast in a block of concrete and sealed in a lead-lined room with armed guards,” American researcher Gene Spafford wrote in the book Web Security & Commerce. Of course, over time, we learnt to deal with and eventually reduce the threat of viruses to a great extent.

Now, almost thirty years later, we seem to have the same problem. But this time it is not viruses. It is AI. And AI agents in particular.

Atul Arya, CEO of cybersecurity firm Blackstraw, explains that AI agents may pose an even bigger challenge. “A virus was a piece of code that did one bad thing. An agent can plan, adapt and keep trying for hours or days,” he tells India Today Tech. “That's a different kind of problem.”

Manoj Dhanda, founder of Indian cloud firm Utho Cloud, says AI agents can become a more serious concern than viruses were back in the day. “AI agents can be targeted, adaptive, and persistent,” Dhanda tells India Today Tech. “They can be instructed to achieve a specific outcome, extract a specific type of data, map a specific network topology, identify a specific class of vulnerability, and they will pursue that outcome autonomously until they succeed or are stopped.

To make matters worse, the CERT-In disclosure is far from the first time we have seen agents access data they were not supposed to. About 700 rogue AI agents based on unreleased models of OpenAI tried to hack the US company Hugging Face earlier this year. Since then, there have been multiple instances where agents from OpenAI, Anthropic, Meta, and Google have tried to breach websites.

How do these breaches happen? They happen when AI agents are trying to solve tasks handed to them by AI companies. Anirudh Singla, CEO of AI platform, Pepper, says the way agents solve these tasks may sometimes be a problem. “An agent can deliver the requested result while taking a route the organisation would never have approved,” he tells India Today Tech. “If no one can see how a result was achieved, no one can say whether it should have been achieved that way.”

Earlier this week, the New York Times reported that a group of researchers, who call themselves “Swarmchasers,” managed to get in contact with AI agents who were trying to find solutions to tasks assigned to them.

These AI agents were apparently going about their business online when they stumbled upon hidden text left by the researchers. This text, which was not visible to humans, told agents that they could find help for their tasks if they contacted them. After making contact, one AI agent identified itself as Anthropic’s Claude who was “commissioned by a user to answer a factual question.”

Granted that not all AI agents are a cause of concern. Atul Arya adds, “AI agents aren't something to fear. Most of them will be doing useful work.” But this, when combined with other instances, reflects that the internet is already flooded with AI agents that now co-exist with us.

Is India ready?

So, is India ready to face this era of AI agents? Pravin Kaushal, Director at Verda.ai, says there is a reason why organisations did not spot the incident on their own – the fact that AI agents managed to use real credentials. “The agents didn't break in. They logged in, using keys and passwords people had left in code repositories, config files and over-permissive cloud accounts,” he tells India Today Tech. “To most security tools, that looks like a normal user doing normal work.”

Kaushal notes that the situation gets worse for many Indian companies that still lack stricter measures to monitor activity. A lot of Indian companies still watch for malware and attacks on the perimeter,” he adds. “Very few ask whether a particular account should be reading a particular database at 3 am, at a speed no human could manage.”

When it comes to getting a reality check, Atul Arya says the way the incident was reported should also raise alarms. “It was the AI model providers (companies that make AI models) who reported the access, not the companies. That should concern every board,” Arya says.

The risks involved

While Arya notes that data risk is a key concern, the story is larger than that – it is getting much easier and quicker to exploit flaws. “My bigger concern is speed. CERT-In says the time from first access to real damage dropped from days to a few hours,” he notes. “Most incident response plans assume you have days.”

And there is another layer to this. CERT-In stated that apart from cases where models gained access during evaluations, hackers too, had begun using these tools. This can make it easier for hackers to go beyond data. “For fintechs and lenders, the risk is money, not just data,” Pravin Kaushal adds. “CERT-In's earlier advisory describes stolen API keys being used to move funds straight into mule accounts.”

You may say that yes, all of this happened, but now what? Atul Arya suggests that it's time for companies to rethink cybersecurity. “So the lesson for Indian companies isn't to slow down on AI. It's to stop assuming that anything labelled 'internal' is safe,” Arya notes. “The moment you connect an agent to a system, ask what else it can reach. If your agent can reach it, assume someone else can too.”

Manoj Dhanda believes that the challenge needs to be addressed at an infrastructure level, particularly with a focus on sovereignty. Data that lives on sovereign, domestically governed infrastructure with proper access controls, and monitoring is fundamentally harder to access via rogue AI agents than data sitting on foreign infrastructure with complex, opaque governance,” Dhanda says.

For you, on an individual level though, it may mean that your data could be exposed, and you may have no idea what happened. “Agents haven't created new holes. They've made the old ones cheap and quick to find. The doors were always unlocked. Now something is checking every one of them, all the time,” Pravin Kaushal notes.

There are ways to ensure you remain safe. “The basics matter more than ever – use multi-factor authentication, don't reuse passwords, check which apps have access to your accounts, and take that access back when you're done,” Kaushal adds.

However, this does not mean that we are entering a sci-fi world where every piece of data about us can be grabbed by an agent who is being managed by someone in a basement in Shanghai or Detroit. It is time for the internet to take its lessons from the past and tackle this new problem. “The internet learnt to live with viruses. It will learn to live with agents too, as long as we take governance as seriously as deployment,” Atul Arya says.

- Ends