The real challenge for CISOs: Unifying cybersecurity silos and data security in the agentic era
Slowed down just to overthink publicly
I’ve always found it easier to think fast than to write slow.
As an operator, speed was oxygen. Decisions needed to be made yesterday. As an investor, context-switching is currency. One minute you’re jamming on how to assess the right hire over a pickle-ball match, the next you're decoding AI infra over coffee.
Writing, though? It demands stillness. Clarity. A willingness to pause long enough to ask: “What do I really think?”
So why am I doing this?
I want to build a community of builders who want to understand how VCs think and of VCs who never forget what it's like to even remotely attempt to build.
I want this to be your tactical unfair advantage and a place where you can hit reply, ask for feedback, or just call out jargon when you see it. This space is for those moments between an elevator conversation and product pivots. The unsexy truths, the frameworks that actually hold up in the wild.
Because you deserve investors and builders who don’t just throw capital but bring curiosity, context, and compound thinking. And because the best ideas often begin as DMs, Twitter threads, or long-form rants that needed a better home.
This is that home. For both of us.
I'm not going to over-promise on frequency (think once or twice a month), but I’m committing to being curious, consistent(ish), and open. I’m kicking off a series of short episodes exploring themes that fascinate me. The goal? Ask the right questions, gather your feedback, and maybe, just maybe, stumble upon an framework to enable builders. Some days we build conversations, other days lego blocks and just once a while - something worth while for you to carry it back when you sleep. Having said that, I am starting off with a mini series on Episodes to decode cybersecurity in the day of Agentic AI (way to go to bridge my personal ambitions with something that has been keeping me professionally curious).
Exploring two key themes: how AI agents can enhance the effectiveness of traditional cybersecurity tools, and the emerging challenges in the age of agentic AI.
CISOs are scratching their brains over unifying cybersecurity silos
CISOs aren’t just focused on firewalls - they’re focused on data and workflows. They want to know: Where is our critical data? Who can access it? How does it move across systems? Visibility into sensitive data stores, API flows, and third-party access is non-negotiable. Equally important is securing workflows. CI/CD pipelines, identity provisioning, and AI deployments - all potential attack vectors if misconfigured. A CISO’s main headache is to ensure every data point and workflow step is monitored, controlled, and resilient against breach, not just logged in a dashboard.
What can AI agents solve for in a traditional cybersecurity framework?
In today’s “AI agent-connected” digital world, the attack surfaces has grown far and beyond. Traditional attack surfaces spanned networks, applications, endpoints, cloud environments, and SaaS platforms. Tools like Snyk (for DevSecOps), Wiz (cloud), CrowdStrike (endpoints), and Proof-point (email) are critical in securing these layers and yes, some of them also cater to Gen AI workflows but they are what I like to call “log generators” - they collect data, highlight risks, and sometimes recommend what to fix. But at the end of the day, they operate in silos. They're great at identifying issues in their specific domains, but none of them truly unify the insights or drive coordinated action across the broader attack surface. Challenge is that an organization ends up building workflows of mitigating risks in silos.
And this worked so far - Cybersecurity as a space had been ridiculously notorious for having point solutions being built out on acronyms that sounded fancy. There was(probably is) a valid reason for this - “Cybersecurity is fragmented by design and necessity”. Every organization has different tech stacks, risks, and compliance needs, which gives rise to point solutions tailored to specific problems: endpoint protection, API security, email threats, cloud misconfigurations, identity access - the list goes on. When this is combined with an ever evolving threat landscape and widening attack surfaces, vendors rush to plug gaps opened by a new type of attack But in reality, these acronyms often repackage capabilities in response to market trends or analyst frameworks (like Gartner Magic Quadrants), not necessarily architectural innovation.
It worked so far, but not with Agentic workflows. To truly understand risk, you need to see the full picture. How agents, systems, identities, data, and assets connect across your environment. Imagine mapping these relationships like a living graph, where one compromised node shows ripple effects across the stack. But none of that is possible without first bringing all your security data into one unified place.
This brings us to the biggest challenge for CISOs - How do you have all of that data together in one place and build insights on top of them all. Next step is building workflows on them all.
The interesting part is who you are targetting - it has to be a security operator-first design. It is not about re-inventing the wheel with respect to traditional security but enabling experts to build no-code/low-code autmation layer on top of existing monitoring systems.
Emerging security challenges
Data security in the Age of Agents
The CISO’s Growing Concern: Securing AI-Native data environments.
Traditionally, data was stored in structured formats, such as tables, which made it easier to implement identity and access management (IAM) policies since each resource (like a table) was clearly defined. However, unstructured data is now being fed into AI pipelines through emails, Slack, cloud apps, and storage buckets, making it difficult to apply traditional IAM policies, as the origin of data is no longer centralized.
Pattern-based and rule-based data classification was also simpler with structured data, as it tends to follow consistent patterns - columns contain similar types of data, and certain identities are known to generate sensitive content. This is not the case with contextual and nuanced unstructured data, which defies such static detection methods.
Similarly, monitoring data in transit was easier with structured systems, as read and write operations followed defined formats and occurred at specific endpoints. In contrast, with agentic AI, data in transit may move through emails, MCP servers, thousands of integrated tools, and even via prompt injection, making oversight far more complex.
Finally, transport logs are significantly easier to generate and monitor in structured environments, compared to the fragmented and diverse sources now involved in AI-native workflows.
Perimter security or ring fencing will not work in AI Agent workflows.
Securing data in the age of AI means three things. One, handle large volumes of unstructured data coming in through various apps/users/prompts. Two, scan fast streaming data classify data on the fly. And three, precision in data classisfication is key.
Model Health and security
The biggest challenge with models is data drift. As prompting techniques evolve, most models become vulnerable, since they are trained on specific attack patterns, even slight behavioral shifts by bots or attackers can bypass security.
Another critical issue is the poisoning of self-learning models through repeated prompt injections, gradually corrupting the model architecture.
To address this, you need tools that can monitor these changes in real time and trigger timely alerts, enabling swift retraining on emerging attack patterns.
That’s a wrap for today. As we move forward, we’ll dive deeper into what cybersecurity and compliance truly mean in an agentic world.