Bold Launches With $40M to Target AI Risks on Endpoints

A startup led by a former Varonis director emerged from stealth with $40 million to better secure enterprise endpoints in the age of artificial intelligence.

See Also: Agentic AI and the Future of Automated Threats

New York-based Bold Security is taking funding from Bessemer, Picture and Red Dot Capital in order to develop AI agents on endpoints, where they can better address AI-driven applications operating directly on user devices, said co-founder and CEO Nati Hazut. Bold aims to offer real-time risk reduction, improved scalability for large enterprises and reduced privacy concerns while eliminating hidden cloud costs.

“I think this piece was neglected for many years by enterprises, and now, because of what we see, we’re aiming at the enterprise endpoint in the age of AI,” Hazut told Information Security Media Group. “Because we do believe it’s fundamentally changed how users are using the endpoint and the risks that we see there.”

Bold, founded in 2024, has been led since inception by Hazut, who established Polyrize in 2018 to map and analyze relationships between users and data across cloud applications services. He sold it to Varonis in October 2020 for $39.4 million. Hazut then spent more than three years as a senior director of cloud solutions at Varonis.

Small Language Models Rather Than Large Ones

Software is increasingly embedding AI capabilities into applications that run directly on the devices. AI assistants, automation tools and local agents can interact with sensitive corporate data at high speed and scale. Existing security tools were not built for AI-powered processes, meaning organizations face risks around data leakage, misuse of internal assets and automation errors.

“Now, we see how endpoints are getting stronger and better and we see a lot of the workload going back to the endpoint,” Hazut said. “It’s definitely getting more interesting to see what’s going on there.”

Instead of large language models hosted in remote infrastructures, Hazut said the company uses small language models minimized, retrained and fine-tuned to run efficiently on endpoint hardware. By reducing model size, refining parameters and tailoring training data to enterprise security contexts, Bold can deploy AI capabilities on user devices without creating massive computational overhead.

“What we took as a challenge here with Bold is to run the AI agents locally on the devices to be able to provide a scalable solution that can also help with real-time risk reduction and not just after-the-fact analysis,” Hazut said. “They don’t have this hidden cloud cost, they don’t have privacy issues there, they don’t have third party risk, they can do real-time prevention.”

Instead of running separate models for each task, the system uses a shared base and swaps out upper layers for different functions, allowing the platform to perform multiple security tasks while minimizing resource consumption. Tools that consume too much CPU, memory, or battery life are often rejected. Bold set a goal of consuming no more than a quarter of the resources used by non-AI competitors.

“It’s not the billions of parameters and getting the privileges of an OpenAI or cloud running out there,” Hazut said. “So you got to be very precise with how you do the things and how you fine tune them. But again, the result is infinite scale and a great product that can serve lots of large enterprises.”

Why Network Monitoring Stumbles With AI Agent Interactions

Many important activities including file manipulation, application usage, and AI agent interactions occur directly on endpoints and may never be visible to network monitoring tools, Hazut said. Endpoint-level monitoring allows firms to observe actions that are otherwise invisible such as which apps users are installing, how internal data is accessed or modified and which accounts are being used within apps.

“We have today several desktop apps that are using certificate pinning,” Hazut said. “If I today want to govern that users are using the right tools, if I want to know which account they’re using, it’s going to be a complete blind spot for me unless I’m on the endpoint. So these type of things, they’re just going to be more and more common.”

Bold’s long-term vision is to create a unified endpoint platform that provides comprehensive visibility into user activity, application behavior, AI agent interactions and data movement. Hazut said this mirrors what occurred in other areas of cybersecurity. In cloud, organizations initially deployed separate tools for different tasks before eventually adopting integrated platforms that combined capabilities.

“It’s not yet another agent,” Hazut said. “You replace other agents, agents that you don’t like with it. And I remember from my previous journeys, people were so dissatisfied from those agents that they were running. But there were a lack of alternatives in the market, not a lot of innovation.”