Glean Tackles Shadow Agents and Enterprise AI Governance
Shadow Agents Raise the Stakes for Enterprise AI
Alan Shimel speaks with Sunil Agrawal, chief security officer at Glean, during Techstrong TV’s Black Hat 2026 coverage. The discussion focuses on enterprise AI governance and the new risks created by agentic AI. Agrawal explains that Glean began as enterprise search, evolved into a ChatGPT-style assistant for internal knowledge and is now focused on helping companies build secure AI agents.
That shift changes the security conversation. Shadow AI was already a challenge for many organizations. Now, Agrawal says enterprises also need to think about shadow agents. Employees may build or run agents without enough oversight. Those agents can access data, invoke tools and take actions faster than traditional governance models can track.
Approved AI Platforms Can Reduce Risk
Agrawal argues that enterprise AI governance starts with giving employees a better approved option. If workers can use a secure internal platform that delivers strong AI results, they have less reason to send proprietary data to public tools. Glean’s approach is to combine enterprise search, retrieval-augmented generation and model flexibility inside a governed platform.
That matters because employees often upload sensitive documents, source code or business data when they ask AI systems for help. In an enterprise AI governance model, that data should remain under company control. Agrawal says Glean helps by using indexed enterprise knowledge, approved model relationships and controls that prevent customer data from being used to train outside models.
The Agentic Harness Needs Security Controls
The conversation also breaks down the idea of an agentic harness. Agrawal compares the LLM to the brain of an agent. The harness gives that brain tools, memory, skills and access to enterprise context. Those pieces make AI more useful, but they also expand the risk surface.
For security teams, the goal is not to block AI adoption. It is to make approved agentic AI safer and easier to use than unmanaged alternatives. That means setting policies around who can build agents, what data they can reach, what tools they can invoke and which actions need human approval.
Security Must Keep Pace With AI Scale
Agrawal also notes that attackers are using AI to shrink the time between vulnerability discovery and exploitation. Defenders need to operate at machine speed and AI scale. That requires visibility, governance and approved platforms that can support productivity without losing control.
For Techstrong TV viewers, the message is practical. Enterprise AI governance is no longer only about chatbots. It now includes agent behavior, tool use, data access and auditability. Glean is positioning its platform as a way to help organizations adopt AI agents while reducing the risks of shadow AI and shadow agents.
Transcript
Hey everyone, it's Alan Shimel. We're back here for our Techstrong TV special coverage of Black Hat. My next guest is Sunil Agrawal.
Sounil's with a company called Glean. But I'm going to let him actually tell us all about it. Let's first welcome Sounil.
Sounil, how you doing? Very good. How are you?
Thank you for coming up here. And we were going to do these on the show floor, but the show floor this year is so crazy and loud, and it was sensory overload. And we've got this great little backdrop that they gave us here in the media center, and so we're doing it here.
I hope you don't mind. It's quiet. We can talk like gentlemen.
Sounil, let's talk a little bit about you, if you don't mind. I'd like to start off with that. Give our audience a sense of your background.
Sure. So I'm currently the chief security officer at Glean. Mm-hmm.
For folks who might not know about Glean, we started as an enterprise search company. We used to call ourselves the Google for enterprise, a single place where you would go and search for all the data that's spread around within your enterprise, whether it be Jira, Slack, Asana, Google Drive, and so on. There was a real need for that.
Exactly. And then when chatbot came about, when ChatGPT was launched in November 2022- Yeah ... we very soon became the ChatGPT for enterprise.
You can ask your questions, get back responses based on enterprise data, not just the internet data, in addition to enterprise. And then, of course, now the big focus for us is providing our agentic platform to our customers so that they can develop all the agents, become the 3x, 10x, more productive, so that they can focus on more the high value things and leave out all the repetitive stuff to the agents. So that's what the company's about.
I have been at the company for about four years, enjoying every bit of it. Been in the security space a good 25-plus years. Worked at your big companies, your Netflix, your Adobes of the world, and now, of course, enjoy my ride at this hot startup called Glean.
Excellent. So I know Glean along probably all those times, and you're right, it was sort of Google of the enterprise. I didn't realize the ChatGPT.
Let me dig in there a little bit. By indexing, because that's what really search is, right? Yes, yes.
Is indexing. So by indexing it all, you just created sort of a chatbot front end that accessed, was it a RAG kind of thing or? It definitely started as a RAG system.
Mm-hmm. Now that's what I would call it. That was the first generation of agents.
Okay. That, hey, what do you do? The LLMs do not have information about your enterprise corpus.
How do you train them- Put there ... bring them the enterprise knowledge, and that's why the RAG architecture, pretty simple. Then slowly, our agentic harnesses became a lot more sophisticated.
Okay, it is not a simple RAG. Now we got to give the LLMs the power to invoke tools, take actions, and that is where we brought in. And you need to do it in an iterative loop.
That's why the agentic looping came about. And now, of course, the third generation of agents is where now you can have agents work in a autonomous manner, okay? So it's no more an assistant, but now it's a true coworker, where it works alongside you, goes on for hours, maybe days, to achieve a task.
So that's where we are in the third iteration of our agentic platform. You know what's amazing? This whole thing, the third iteration, you think we're talking years, we're talking weeks and months.
Exactly. This is the world we live in now. Absolutely.
What do you mean by third iteration? We had our second iteration yesterday, and the first iteration was last week. But it's crazy.
That's crazy. Crazy. Absolutely.
Crazy. So I got to ask a silly question. What's a company like Glean doing in a place like Black Hat?
Yeah, so security is actually, because we deal with so much data and so much autonomous agents, security is right front and center for us. Yeah. It is one of the core pillars for what we offer to our customers.
Trust, data governance, AI governance is right front and center. Hence, having presence at Black Hat for us is very important, not just for making sure that all the agentic platform that we provide to our customers is secure from ground up. At the same time, in this post-Mythos world, we are also thinking about how do we protect our company?
Because now your attackers have all these powerful tools, and the time to window from vulnerability to exploit, which was actually over a year in 2021- Right ... has come down to a few days. And if- If not hours.
Exactly. And if this continues, it would be in minutes by the end of next year. No, the whole thing is upside down.
Exactly. It's changed. And the attackers have access to those tools.
Now, the defenders also need to act on machine speed, and that's where I'm here looking at one of the promising companies that I should be a part of. Absolutely. And I got to say, so it's not just machine speed.
So I call it AI scale. It's speed and breadth. Absolutely.
And that's the thing. Look, if I had to run on this line, I could run on this line as fast as I can. If I've got to run on a mile wide and that way- Yes ...
that's twice as hard. Sounil, there's another issue, though, that I want to bring up, and that is this whole, we used to call it shadow IT. When the cloud came out- Right ...
people would whip out their credit card, they'd put a couple instances up. Who the heck knew what was going on? Now we have a shadow AI problem.
Everybody's using AI. Everybody's putting agents loose, who we don't know what agents are out here, what's going on. We have it.
Techstrong's a small company. We're part of Futurum Group. We put out an AI policy, not very sophisticated, but I know in my gut, we really don't know who's using what, what's running loose, what IP do we have, has been put into these frontier models.
Is this something Glean can help with? Yeah. So I would say two parts.
One is, we are the recommended solution. So, when enterprises adopt Glean, you are giving your employees the same power that they would get, whether they were using Claude or ChatGPT, or even more than that. So now this is a officially approved application which has all the functionality that all the other public versions have, plus has the enterprise knowledge.
So now the motivation for your employees to go and try out something else which is unapproved has dramatically reduced. Why? Because they're getting better responses, better productivity gain by using the officially approved application.
So that's one prime way that Glean helps. And the second part is, you talked about shadow AI. There's shadow AI, there's shadow agents.
Shadow AI is so 2024. Nowadays, it's all about shadow agents. It's all about that, yeah.
People are developing agents, and those agents are running wild. Yeah. Correct?
Again, by adopting an officially approved platform like Glean, which has security and governance built in. Into it. Not only you're bringing productivity, you're bringing it in a safe, secure manner, and at the same time, being very usable.
Got it. So you blew my mind with the shadow AI 2024. I thought I was finally caught up.
But you're right, it is. This is the new world we find ourselves in with these agents. And again, even my company, which is a small company, this is our life right now.
Yep. This is who we are. We hear a lot about the term like harness, right?
You could put a harness on it. We're gonna put a harness on it. Is this part of the harness?
Yes. So always think about LLM. That's the brain.
Right. And now you've got to give the brain the hands, the legs, the fingers, the torso. That's what the harness is, correct?
Because the LLM, it's only trained on a certain amount of data. But if you need to bring enterprise data, how do you do that? We talked about RAG.
Right. Another way is by tool invocation. Okay.
Now, at the same time, you've got to give it memory. Right. Correct?
That, hey, Sounil interacts a certain way, and Sounil has done certain actions in the past. How do you do that? You bring in memory.
Right. But then there's certain things that Sounil always looks out for, creates a PDF or a PowerPoint presentation a certain way. How do you codify that?
You codify that in a skill. So these are all the ways that you are giving an arm and a leg to an amazingly powerful LLM, which is the brain of an agent. Love it.
I love it. All right. I got to pivot us a little bit now.
So you said Glean can help with this. Let's get some specifics down. How does Glean help with this?
Yeah. So what Glean does, one question you said, someone might want to use ChatGPT because they have a question. Correct?
And in order to have the question, they might upload a company proprietary IP or some company's source code because they want to fix a bug or ask a question on a PDF. Right there, what has happened is your company IP has leaked to a public version and where they might train on the data. Instead, if you are using Glean, all that data remains within your control.
You're getting the same functionality or even better functionality because we are not only agnostic of any large language model. So we get the power of the GPT models, the Claude models, the open weight models, everything combined. Plus, the data is already indexed.
You don't have to upload the PDF. Agreed. You can just go ahead and ask a question.
If it has to reference data from 10 different documents or your Slack conversation or your email, it knows all of that. So not only are you getting a better answer, you're getting it much more effective because we will use N number of models, which is the best model for the given subtask. You're getting it cheaper because if a certain model we can use, one of the open weight models, we will use that, and it's all secure and governable because all the data always remains within your control, and none of the models, because we have contractual relationship, that they cannot train on any of your data.
So there are multiple advantages to using an official platform like Glean. Tell our audience, how can they get Glean? Yes, absolutely.
com, or just go on our website and request a demo, and one of the helpful Glean employees will reach out to you and make sure that all the power of AI and agentic AI can be brought to your enterprise. I love it. Sunil, thank you so much.
I appreciate you coming on here. I know you're busy running around. It's crazy.
Yeah. Have you been to Black Hat? How many times you've been to Black Hat?
Oh, I've been coming here for far too long. Have you ever seen it this crazy? No, this is actually, I would say definitely, as you said, it's a sensory overload.
But at the same time- It's kinetic energy, though. Exactly. Correct.
So that is, I wouldn't have it any other way. I love it. We'll be back in touch.
Good luck. Sure. Thank you.
Hey, we're here at Black Hat. You're watching Techstrong TV.
