You have a background in cryptography. Is the pace of adopting post-quantum cryptography [which secures systems against attacks by quantum computers] satisfactory around the world, considering quantum computing appears closer than ever to threatening today’s encryption standards?
We at Google regularly track the state of progress [in quantum]. It’s a field that’s not linear, so sometimes there are major break throughs. We have some brilliant mathematicians who are on top of this space. What we did just recently was to actually pull the deadline for adoption [of post-quantum cryptography] to 2029. We are trying to set an example for everybody and we are warning them that the time has come for you to migrate your systems. The other application, something Google is not directly involved in, but is going to be problematic [in a world where quantum computers can crack the maths underpinning today’s encryption standards], is the Bitcoin protocol [which is underpinned by cryptographic math]. So that might be another incentive for people to build things faster and prepare for a postquantum world. Looking at all of those factors, the world needs to move faster on postquantum cryptography.
The 2029 date what were the signals that led Google to suggest that?
It’s all the research that is happening.
But countries like China which have invested heavily in quantum wouldn’t want to disclose the state of their progress, right?
Yeah, it could be, and you’ve seen it in maybe the AI space as well. But when you look at the frontiers of research in any topic, especially when you have brilliant minds everywhere working on something, the gaps are a matter of one year or six months. I mean, it has be rare that someone has like a two or three year advantage, and the rest of the world is completely unaware.
Also read | India’s IT model is out of date. Can the industry reinvent itself?
Humans were supposed to be the weakest link in cybersecurity. Has that changed now that AI is being used actively by criminals to find vulnerabilities?
What we see coming more from our threat intelligence team is that humans used to be the vector through spear phishing [highly targeted attack]. That is now maybe slightly less than 50%. What has come up are the number of zero-day vulnerabilities attacks that are exploiting a vulnerability [enabled by AI deployment by bad actors].
In an agentic era, how much will user education on its vulnerabilities matter?
Just like in other areas, the user should be aware of the limitations of the agents.
In India, tech has created a new set of vulner abilities like digital arrest scams. When so much agency is given to algorithms, how much thought goes into designing guardrails, especially for vulnerable communities like the senior population?
This is definitely something that’s top of mind for us. We have to think about the populations at risk. Children, too, right? We can put guardrails and controls on what an agent can do. So we put those controls and build them in. It’s again more of a partnership between Google and other companies in not just building controls but also in educating the population. It’s like, you know, my mother used to say, don’t keep your jewellery at home, keep it in the locker. It is the equivalent of that. So if someone comes to my house and steals my locker key, they still have to go to that bank and prove their identity.
Google has a limit of Rs 10,000 or $100 for agentic AI transactions. How did you arrive at that number?
As of today’s technology and where we are today, you still have threats against agents. Everybody knows about prompt injection [which targets LLMs] and so on. The other piece that people don’t talk about as much is that some times you don’t need to even have an attack. The agents, in their effort to please you, will do things that you may or may not agree with. There are reasons why agents, in some environments, could be unreliable. So you want to have this balance between enabling the value that you get from having an agentic system, but also have some kind of guardrails and some control. We always talk about having the human in the loop, and it matters. So that’s where guardrails come in. I think the thought process was okay, under $100 or Rs 10,000. It seems to be an amount that is okay for a pre approval limit
What are some of the biggest generative AI security risks that enterprises are missing?
Security is no longer a human scale thing, it is machine scale the speed at which attacks are happening, the fact that vulnerabilities will be found quickly [by bad actors].
Also read | An ambitious project’s smart city dream runs into major hurdles
Have you looked at your enterprise from a perspective of what happens when AI attackers are that much more powerful? Sometimes it’s just having the capability to react fast, the bandwidth to quickly contain it if there is an incident. That is a test for everybody.
CEOs often wonder what is the right mix of tools to have in cybersecurity. In the AI era, do you see more consolidation of tools to make the lives of executives easier?
I think you’re right. There was a time when you had to assemble a whole bunch of tools. Now we are seeing customers asking for more integrated defences. I think this may also be driven by the agentic era. You need to have automation, and automation means you have to have a bigger picture.
Sometimes you don’t need to even have an attack. The agents, in their effort to please you, will do things that you may or may not agree with”