What are the key cybersecurity themes emerging for enterprises in an age where artificial intelligence (AI) can potentially exploit vulnerabilities?
If you look at cybersecurity from a macro perspective, there is a paradigm shift that's happening. Be it for consumers, enterprises, or governments, all of them have been affected by cyber attacks in the past. In the past three years, AI has gradually crept in and there is a situation where the attack surface and defence surface are both enabled by AI. Entry barriers for an attacker have reduced. An attacker can get an AI package and launch phishing attacks on users, which applies to fake audio and video calls, where the entire spectrum is getting covered by these multi-modular large language models (LLMs).
How is it different today than the previous years?
For the last two decades, enterprises have been deploying what was called state-of-the-art security in data leakage prevention systems or cloud security posture management, including other expensive solutions. Today, the problem is where an individual is passing confidential company information to the internet via any popular LLMs. Confidential information is likely being trained by the LLM provider. This is called Shadow AI (use of AI tools, applications, or models by employees without the approval, oversight, or visibility of an organization's IT and security teams). We are solving customer problems around that. This includes getting solutions from the market, educating enterprises about this issue, enabling proof of concepts, doing a gap assessment, and showing how their confidential information is going out, and then getting them to deploy the solution in their network.
Do you think the answer to Shadow AI is enterprises building small models themselves for internal use?
Not everyone can afford to run their small language models (SLMs) or LLMs. What we do there is look at the security portion of the company, their roadmap towards certain plans. If they want to deploy their own model, what are inherent security challenges? How clean is the data pipeline that you use to train your internal AI? Is that being injected with any malicious data? How are you creating your prompt engineering structure? One of the services that's very important an enterprise takes up is how do you do a very clean deployment. It is important to ensure identity management, not only for humans but also agents need identification.
How do you establish identities for AI agents?
We are creating an infrastructure for establishing digital identity for non-human agents. That includes managing that life cycle of that identity carrying every data that that AI agent handles if it goes rogue. It also includes compliance, usage of tokens, metering their usage, among other things. Also, focus on what are the policies that need to be implemented, and which platform will be used to implement.
What do enterprises expect when it comes to non-human agents’ compliance?
For non-human agents, there are proofs-of-concept that enterprises have deployed. We provide services to plan, architect, develop, and implement the project. Enterprises want to restrict open usage by agents. We are focused on what happens if an agent goes rogue, what are the compliance checks involved, token and identity management, and eventually the entire spectrum of the life cycle of non-human agents.
 "Srinivas L, joint managing director and chief executive officer at 63SATS Cybertech")