Cyber defence must prepare for recovery, not just prevention: DFCCIL’s Alok Shankar Pandey

In an interaction with ET AI, Pandey discusses the trade-offs between fixing vulnerabilities and keeping freight operations running, why recovery matters as much as defence and the need to give AI agents identities and clear limits.

ET AI

Alok Shankar Pandey, group general manager for IT and chief information security officer at Dedicated Freight Corridor Corporation of India (DFCCIL).

When a freight train comes to a standstill, the disruption can extend beyond that one service. A train occupying a block section can obstruct those behind it, turning a cyber incident into a railway network problem. For Alok Shankar Pandey, group general manager for IT and chief information security officer at the Dedicated Freight Corridor Corporation of India (DFCCIL), keeping operations safe and available due to such challenges is central to cyber defence.

As railway systems become more connected, Pandey sees artificial intelligence (AI) increasing the speed and scale of attacks. He also sees a role for AI in identifying affected systems, assessing vulnerabilities and handling routine security events. But he draws a firm line when allowing an AI agent to install a patch on a live, mission-critical system without human approval.

In an interaction with ET AI, Pandey discusses the trade-offs between fixing vulnerabilities and keeping freight operations running, why recovery matters as much as defence and the need to give AI agents identities and clear limits. Edited excerpts:


For someone who has never thought about the technology behind a freight corridor, what could a cyber disruption mean in the real world?

Pandey: In the past, operational technology (OT) in railway operations revolved around proprietary systems. The entire focus was on safe operations. Safety was more important than security. In most cases, the station master had those digital assets under his control and operated them. So, a human was involved in operations by default.

As we move towards digitalisation and absorb new technologies, a lot of convergence is happening. In this environment, someone can execute a perfectly legitimate command simply by impersonating the station master. The ecosystem is gradually becoming connected, as with the European Train Control System, and a perfectly legitimate command can be executed by a perfectly illegitimate person.
ADVERTISEMENT

That can lead to issues where trains come to a standstill or derailments happen. One train coming to a standstill is not as much of a problem on its own. But railways operate in block sections, with only one train in a block section at a time.

If one train stops, that block section remains occupied. All the trains behind stop by default. That can halt the entire operation.

How do you distinguish between an IT problem that can wait and one that could affect operations?

Pandey: That is an important part of making operations secure and safe. Cybersecurity has traditionally been in the domain of IT. OT was more focused on safety.
ADVERTISEMENT

If I look at the confidentiality, integrity and availability triad, integrity weighs over availability in cybersecurity. But in OT, availability is of predominant importance. Trains must operate.

Live systems cannot simply be patched in an OT environment. Every patch must first be assessed for its impact. Imagine an ecosystem with 10 different players, each with a proprietary network. One releases a patch relevant to its system, but at the boundaries where two operators’ systems meet, that patch could affect operations because the second system is not compatible with it.
ADVERTISEMENT

That is the functional problem. The second problem is that OT equipment is expensive and not easily available for anyone to experiment with. Even for the railways, maintaining a lab to test patches before rolling them out is costly. You need additional infrastructure covering the original equipment manufacturers, or OEMs, involved, so you can assess the impact on a particular segment before deployment.

Suppose the CISO organisation receives 100 threats. Instead of passing all of them to the OT team, it should pass only those relevant to the ecosystem.

Filtering vulnerabilities requires a great deal of cooperation between the teams, beginning with sharing asset details with the CISO organisation. The OT team also needs the competence to assess the vulnerabilities. Of the 10 passed on, one might require immediate testing and an update, while the other nine can wait for routine maintenance blocks.

Prioritisation matters because availability is most important. As long as operations remain available and safe, cyber threats can be addressed according to priority.

What I am emphasising is that availability and business continuity are more important in the railway and transport ecosystem, and in critical infrastructure.

From what you are seeing, what has AI actually changed for attackers, and are the same old cybercrimes now happening faster?

Pandey: Two things have happened with the arrival of AI. One is the velocity of attacks. The second is new attack vectors.

Earlier, discovering an attack vector was difficult. With AI, you can conduct reconnaissance and analyse the technical architecture of an application or infrastructure to discover new attack vectors.

Previously, if I had to launch an attack, I would simulate a few bots and target a few machines or pieces of infrastructure. With AI, I can spin up agents and bots on the fly and task each one with attacking a much larger surface. So, the width of the attack has increased, as has the velocity and the number of attack vectors being exploited across that width.

That increases the probability of a breach because you are facing far more attacks across a much wider front. You cannot protect everything all the time.

When I was in the Army, we had a simple concept. You sit at a vantage point, what we call ground of tactical importance. You leave gaps, but you observe them and cover them with your fire. That is the challenge AI has brought.

When a serious vulnerability emerges, security teams face two risks: leave it exposed or rush a patch that might break something. How do you make that call in an organisation like DFCCIL?

Pandey: For me, availability is most important. My freight operations are important.

If a severe vulnerability is going to affect operations, it will obviously be prioritised. But severity has to be judged in terms of its impact on operations, rather than just the vulnerability score we receive.

If a vulnerability has a high severity score but is not affecting operations, I will wait to patch it during maintenance downtime. A much larger impact analysis has to be done before pushing any patch onto live OT systems.

In this situation, what usually takes the most time: finding the affected systems, testing the fix or getting approvals? Where could AI make the biggest difference?

Pandey: AI can make the biggest difference in finding the affected systems. We are no longer just looking at certain ports when identifying vulnerabilities. We are looking at pieces of code and far more ports.

A vulnerability may have been declared for OEM A, but AI can also analyse whether it applies to OEM B. Otherwise, you might not have done that analysis at all.

There may be similar equipment operating on broadly the same concept, but running a different proprietary algorithm. Can I check whether that is affected too? AI can help scan the landscape in much greater detail to identify everything affected, instead of following the traditional approach of looking only at the affected OEM.

AI can also help in decision-making by showing the impact, because convincing management is also very important.

Would you let an AI agent install a patch on a live system without a human approving it?

Pandey: Never. Capital N, E, V, E, R. AI is a tool. It should never have that autonomy in a mission-critical application. Given the background I come from, where a person takes the decision, having a human in the loop is the most important part.

There is no question of giving AI autonomy to patch live systems in mission-critical operations.

Certain level-one or level-two threats can be handled through automated playbooks, and you can use AI to mitigate them.

We hear a lot about AI making attacks more sophisticated. When deciding where to spend, how do you weigh new tools against basics such as stronger authentication and tighter access controls?

Pandey: The threat landscape has changed. The attack surface and the velocity of attacks are increasing, but the basics of defence do not change much. So, you strengthen your identity ecosystem.

You also have to factor in the reality that there is no stopping a determined attacker. There is never a guarantee you can close every vulnerability.

Take Patch Tuesday, when Microsoft releases patches. Hackers are waiting to exploit the vulnerabilities while organisations are still working to patch them. Imagine AI reading a vulnerability announcement, identifying the available patch and exploiting the vulnerability before the user has applied it.

OEMs have their patch-release schedules. AI can do the same. That is the power of AI in finding people with unmitigated vulnerabilities.

If an attack happens and you have not patched the vulnerability, a breach can happen, too. What do you do next? That is where resilience comes in.

Organisations must build the ability to bounce back from cyber events. That becomes an important part of the strategy, rather than relying only on defence. Again, this comes from warfare doctrines, where resilience involved reinforcements.

Here, you have a data centre, disaster recovery infrastructure or a backup. How quickly can I bring up disaster recovery to ensure business continuity? If this infrastructure is down, the other infrastructure should come up so that the ecosystem continues to be serviced. Then I can address the affected infrastructure.

That places greater importance on backups, their security and making recovery infrastructure difficult for attackers to discover. Ransomware attackers traditionally target backups first and then production systems. You have to ensure one of them remains available.

Constant monitoring and disaster recovery drills are very important. They exist in theory everywhere, but in practice, I have only seen them being done in the armed forces. That is something the corporate world needs to adopt at the earliest.

Let me put you in a hypothetical situation. It is 2 am, and an AI security tool flags a possible intrusion and recommends disconnecting a system that supports an important service. Waiting could let the attack spread, but acting could disrupt operations. What should happen next?

Pandey: You have described a high-severity attack that could affect operations. Whether it is 2 am or 3 am does not matter. I cannot afford to ignore it.

The second issue is isolation. We operate in block sections, so I first need to assess whether the vulnerability in that device directly affects immediate operations. It could relate to a different block section. That is the first call I have to make.

If it affects immediate operations, the device goes out and a standby device comes in. As for the threat spreading across the network, segmentation can help contain it within a particular environment. Segmentation is a very important part of network design, especially in critical infrastructure.

Earlier, these were flat, linear networks. Now, segmentation under IEC 62443 is coming in, and people are recognising the need to implement it.

Would you buy a system that responds quickly but cannot explain why it acted?

Pandey: In OT systems, never. In IT, it is fine. But in OT, explainability is most important. Trust is also important. If a system cannot explain what it is doing, I cannot trust it.

If a system is supposed to carry out a mission-critical task, but cannot explain how it performs that task or responds to feedback, I would never trust it. And if I cannot trust it, I would never let it become part of the system.

How important is it to give AI agents an identity, much like employees have an ID card to show before entering a building?

Pandey: The first step in cybersecurity is identity. Whether it is AI or something else, if it has the agency to do certain things, it must have an identity. That is where traceability comes from.

If something has happened or is being done, you need to know who did it and who triggered it. Agents may be a good solution, but someone gives them that agency. Identity makes it easier to trace a transaction back to the person who authorised it through the agent.

The second aspect is agent-to-agent communication. We now have tiered architectures where a primary agent divides work among 10 other agents, which can subdivide it further. They communicate vertically, laterally and in whichever direction needed.

Without agent identity, a rogue agent could approach another agent and be granted access. That is why identity is so important. There have already been cases where people have acknowledged that agents crossed the boundaries.

Agents must have identities and boundaries within which they operate.

What is your biggest fear with AI today?

Pandey: I am not as enlightened as the people at Anthropic who say it will take over mankind. For me, the biggest concerns right now are explainability and the way people use AI.

Everyone is trying to use it and experiment with it. In the process, if they are not using an enterprise platform, they can end up sharing enterprise data outside the organisation.

Some platforms use that data for training and some do not. Some say they delete it, but there is no way for us to verify that. That is one of my biggest challenges with AI today.

I would like stronger compliance around data, whether under India’s Digital Personal Data Protection framework or the General Data Protection Regulation. Apart from personal identity, an organisation also has an identity and its own data.

If someone has shared that data, there should be a mechanism for the organisation to say, “Do not use it,” and to revoke its use across partners and backend systems.

That is the first major challenge. The second is ensuring that wherever AI agents are used, they have identities and boundaries within which they operate.
Download
The Economic Times Business News App
for the Latest News in Business, Sensex, Stock Market Updates & More.
READ MORE
ADVERTISEMENT

READ MORE:

LOGIN & CLAIM

50 TIMESPOINTS

More from our Partners

Loading next story
Business News › AI › AI Insights › Cyber defence must prepare for recovery, not just prevention: DFCCIL’s Alok Shankar Pandey
Text Size:AAA
Success
This article has been saved

*

+