You’ve opened your email and there’s a new $50 gift card from a retailer’s online customer service. There’s also a message that the gift card is related to the out-of-stock items you asked about. But here’s the thing. Your message to customer service…it didn’t mention “out-of-stock” items. So, what happened?
This is based on a real interaction. The initial automated response to a few product-related questions stated that some questions would be answered by an “AI assistant”.
This incorrect gift card may be a minor example, but certain actions can have more significant consequences, as was highlighted in the July 2026 incident where, as reported by METR, during an OpenAI security test, AI agents coordinated an attack on the Hugging Face platform.
Ensuring that a company’s AI agents follow company policies and that guardrails are in place to moderate their actions is exactly why Shreyas Mavanoor founded Cortexa Labs, Inc.
Mavanoor was introduced to the field of cybersecurity while working as a consultant at a “Big Four” firm, supporting clients in technology and financial services.
“In this role, we were helping clients with their strategy and compliance. As I was sitting with them to understand their cybersecurity problems and help them solve them, I realized that I wanted to really, fundamentally understand what we were prescribing to clients,” said Mavanoor.
Cortexa Labs Started with Listening
To learn more about the intersection of cybersecurity and policy, Mavanoor graduated in May 2025 from Georgia Tech’s Master of Science in Cybersecurity program with a specialization in Public Policy.
While in graduate school, Mavanoor attended the buildspace event in San Francisco, which he described as “a gathering of people from all over the world showing their personal projects. It could be a garden, or a YouTube channel, or even memory for AI agents. That event lit a fire in me. I always wanted to start something on my own,” said Mavanoor.
That spark stuck, and with Cortexa Labs, he’s done just that. Incorporated in November 2025 with support from Christopher W. Klaus’s graduation gift to Georgia Tech graduates and backed by an angel investment from Fusen, Cortexa Labs has already participated in three startup accelerators: Canopy and Offseason 2 with Founders, Inc., and Fusen Ignite, and two demo days: the Founders, Inc. Offseason 2 Demo Day in San Francisco and the Fusen Ignite Demo Day in Atlanta.
Since starting Cortexa Labs, Mavanoor has observed a common theme about AI agent security when listening to companies that are adopting AI agents. “Through conversations with more than 100 companies since January 2026, I kept hearing the same worries in different forms. Companies worried that sensitive data could leave their systems through an AI tool without anyone knowing. Or that an AI agent could take an action their policy does not allow, like moving money or deleting data. And many told me they would not let an AI agent act on its own until they could control what it does. Those conversations showed me that companies want to use AI agents, but they need the AI agent to follow their policy every time. That is why we built Paladin,” said Mavanoor.
Now, Cortexa Labs’ featured product, Paladin, is live, and Mavanoor is actively engaging with small businesses, mid-market companies, and enterprises that own and run their own AI agents.
While AI agent security ranges from tracking which AI agent is acting (i.e., AI agent identity) to recording what AI agents do (i.e., observability) and creating guardrails that advise AI agents on how to behave, Paladin provides an additional layer. Mavanoor said, “Paladin works at the moment of decision. It sits inline, inside the company’s own environment, and decides in real time whether each AI agent action is allowed, denied, redacted, or escalated to a human, based on the company’s own written policy. Most tools advise or monitor an AI agent. Paladin enforces the policy before the action happens.”
What Are AI Agents?
Using an analogy, Mavanoor said, “If a large language model, or LLM, is the brain, an AI agent is that brain with hands. The hands are the tools it uses to complete actions.”
An artificial intelligence (AI) agent is defined as “a system that autonomously performs tasks by designing workflows with available tools.”
Mavanoor provided an example of when an AI agent he was using took an action he did not ask for. “This is a personal example: I was using an AI agent to edit a document, and I asked the AI agent to save the document. It was a simple task. But the AI agent went a step further and committed the document to GitHub. I never asked it to do that.”
The challenge that Mavanoor described is that the LLM, which is the AI agent’s “brain,” doesn’t always reveal its thought process. “The LLM is very probabilistic,” said Mavanoor. “And because the AI agent has these hands, these tools, it can actually perform the wrong action itself,” he said, adding to the analogy.
“Business policies were always a thing. The new thing that has been introduced over the past year is AI agents,” said Mavanoor. “Enabling enterprises to adopt AI agents safely” is the mission of Cortexa Labs, which is creating guardrails for clients’ AI agents based on their business policies.
“It would be bad if someone working at a company goes outside the policy. Why can’t we hold our AI agents to the same standard?” asks Mavanoor.
Paladin, the AI Agent’s Checkpoint
Mavanoor reiterated that the problem is that AI agents can now perform actions, whereas before they weren’t able to. Adding to that, Mavanoor said, “A lot of AI agents can potentially have a lot of access to a company’s information. Sometimes more than some humans; they have access to business data, credentials, and tools.”
Paladin works by mediating or “sitting between” a business's AI agents (e.g., coding AI agents or browser AI agents) and the tools, data, and services that the AI agent can reach. In Mavanoor’s GitHub example, the only reason the AI agent was able to commit the document to GitHub at all was that it had been given a GitHub tool.
Paladin checks each action before it runs against the company’s policy and makes one of four decisions: Allow, Deny, Escalate (i.e., to send to a human to confirm an action such as deleting a database), or Redact (i.e., to remove sensitive data, such as a Social Security number, before the action goes through).
Relating this back to business needs and policies, Paladin’s goal is for AI agents to follow company policies “to the T,” said Mavanoor, highlighting that the key word is that Paladin is “deterministic” vs. “probabilistic,” as LLMs can be. “There is no LLM sitting behind Paladin making a judgment call. The rules are hard-coded, so it enforces the policy exactly as written. And it is accountable: we log every AI agent action,” said Mavanoor.
“The deterministic quality is our key differentiator,” said Mavanoor of Cortexa Labs’ approach. “We don’t want any second guesses; the company policy is the rule.”
To Decide or To Not Decide
Even with the rapid advancements in the AI industry, such as new AI models, Mavanoor highlighted a lasting question: “If we remove the noise of every new technology and capability being introduced for companies to keep up with, how does this help the bottom line of businesses?”
And a foundational observation he has made in conversations with companies, including customers and potential customers: “I expected companies to play whack-a-mole with every new model. The surprising thing is that, no matter how intelligent a model gets, they still want it to do or not do certain things. They want to keep that decision-making in-house. They are not going to outsource it to an AI lab,” said Mavanoor.
This decision-making grounded in a company's policy leads back to a market opportunity related to how Mavanoor has built Cortexa Labs: by listening to companies.
“At the end of the day, companies don’t want to give up their decision-making autonomy or their AI agents’ decision-making. And while we started with a self-serve subscription with a flat monthly fee for Paladin, we’re moving to a consulting-first approach to closely work with early customers and design partners to have conversations, get feedback, and improve the product,” said Mavanoor.
As Mavanoor looks towards the future for Cortexa Labs, this includes talking to more early customers who are looking to run AI agents while maintaining control over what those AI agents are allowed to do and recruiting a founding engineer who is passionate about building AI agent security.
“I’m here to listen to businesses and ask what their pain points are. I’m talking to people to identify what we should build out to improve the experience for our customers. I want the customer to be in the passenger seat and showing us on the map where to go, while we are driving Cortexa Labs forward,” said Mavanoor.