The AI Securing Your Workforce Before a Mistake Ever Becomes a Breach | E2314

Brandon Dixon is the founder of Int, a startup that emerged from stealth with a $100 million raise focused on preventing corporate policy violations before they escalate into security breaches. Dixon argues the security industry has largely abandoned prevention in favor of reactive detection, and that roughly ninety-nine percent of insider risk cases involve mistakes by well-intentioned employees rather than malicious actors.

Int deploys an on-device agent on endpoints such as laptops and phones that assesses in real time whether a user is about to violate corporate policy and intervenes before the action is completed. Sub-second decision speed is a hard technical requirement, and the system uses embedding models specifically because they are performant enough to run on CPUs without GPUs and can return decisions within that window. Dixon credits recent advances in embeddings, which translate semantic meaning into representations computers can process, as the capability that made this feasible now when it was not possible a few years ago.

The product monitors human behavior, interactions between users and AI agents, and the behavior of AI agents themselves. Dixon identifies AI adoption inside enterprises as a growing risk surface, pointing to citizen developers encouraged by leadership to use AI tools without technical backgrounds, who risk accidentally pulling sensitive context into those tools and leaking it outside the corporation. Non-AI mistakes such as emailing financial data to the wrong recipient or sharing credentials in chat systems are also within scope. Int builds continuous behavioral baselines for individual users, departments, and peer cohorts over an initial observation period of approximately two weeks before shifting into intervention mode, and Dixon says the product requires minimal setup, with something as basic as a list of sanctioned software sufficient as a starting point.

The product is deployed entirely within the customer's boundary, meaning Int itself does not see customer data. Global 2000 and above enterprises are the target segment, and self-hosting is offered as a one-click deploy inside the customer's chosen major cloud provider, with lower licensing costs for customers who absorb their own hosting expenses. Dixon distinguishes Int from traditional EDR vendors, describing EDR as a commodity at this point with most enterprises already having a solution in place, and positions Int as augmenting where EDR, insider risk, and DLP tools fall short rather than competing directly with them.

The behavioral data the product collects for security purposes carries secondary value. It can accelerate closure of true positive benign tickets in the SOC, provide additional context to DLP events, surface insider risk activity, and identify employees who need training. Dixon also suggested the accumulated behavioral substrate could identify mundane repetitive work across departments that AI agents could take over, a use case the hosts predicted could extend the product well beyond cybersecurity into operational efficiency. The hosts further predicted that AI labs, joint ventures, and private equity groups may eventually attempt to acquire the company specifically to access its behavioral data. The company has hired more than 75 people without paying a single recruiting fee, and current hiring is focused on sales reps, sales engineers, and forward-deployed engineers to handle inbound demand Dixon described as a floodgate of interest following the stealth announcement.

David Im, founder of Sumay Labs, presented an agent orchestration layer that wraps five or six video generation models alongside image, audio, and clipping models into a single API, with the goal of generating high-quality video outputs in one attempt rather than through repeated iterations. The core technical challenge is that video generation models are stochastic, making results unpredictable and requiring multiple prompting attempts to achieve usable output. Current models support a maximum of 15 seconds per clip, with newer models extending that to 30 seconds, and Sumay's system routes across multiple models and stitches results together with a single audio file to produce up to 60 seconds of consistent video, though the host characterized this approach as something of a hack rather than a native capability. Maintaining facial consistency across the full video duration is a key technical problem the product is designed to solve.

The user experience requires only selecting an avatar and writing a script, with scene generation handled automatically. The product currently has 20,000 users, with approximately 90 percent of paid customers being brands concentrated on Instagram and TikTok. Im's stated six-month goal is to build a video agent capable of one-shotting full marketing videos for brands and marketers.