The Browser Isolation Product Behind Chrome’s Incognito Window | Bernard Harguindeguy

What we talked about.

In this episode of the Sunny Ray Show, host Sunny Ray talks with Bernard Harguindeguy, a security founder turned investor. Bernard describes the browser isolation product his team built and sold to Google, which aimed to keep malicious content away from a user's machine through kernel level isolation and included a protected banking session and an incognito window. He traces his path from satellite and networking encryption work to enterprise security, then explains what he brought as a chairman and why Elastic Beam, founded in 2014, had to keep machine learning costs low. Bernard recounts selling Elastic Beam to Ping, becoming its CTO, and taking the company public. He shares his view that founders should build partnerships that can grow into acquisition offers, and that good investors support CEOs through the downs. He also discusses his cybersecurity and AI fund, why domain expertise matters in a founding team, and why rule-based controls are the right approach for AI agents today. The conversation closes on why so many enterprise agent pilots fail to reach production.

A founder who sold browser isolation tech to Google explains how to build a sellable security company and why AI agents need rules.

The questions, and the answers.

What problem were you trying to solve with the browser isolation product you built and sold to Google?

We wanted a very secure browser environment where you could open any site or download any spreadsheet without infecting your machine. That required complete isolation at the kernel level. We also protected the reverse case, giving you an isolated banking session even if your machine was already compromised. The incognito window was already part of the product.

What did you see in browser isolation that others missed?

I'm sure others aspired to build a very safe browser experience too. The difference was that we committed to isolating the browser completely from the machine, which means kernel level isolation. It is incredibly difficult technically, and we had a great team. That team went on to drive a lot of Chrome development within Google after the transaction.

What can a chairman do that a CEO cannot?

A CEO is heads down in the weeds most of the time, which is a difficult job. A chairman brings experience about what has and has not worked, plus relationships and contacts. A good chairman also helps the team understand the future and where to go next, because you cannot stay a one product company.

What was hard about AI for security when you built Elastic Beam, and what is still hard?

Back in 2014, the challenge was keeping the cost of machine learning low enough to make sense for customers, since your cost cannot outrun your product's benefit. We innovated to do sophisticated modeling without a huge compute bill. What is still hard is modeling when multiple variables change at the same time. The more variables, the harder it gets.

What do founders systematically get wrong about making a company sellable?

Many founders wait until they want to cash out and then push the company to potential acquirers, who must do a lot of internal analysis. The better route is to build partnerships as you grow. Partners already know your company, where it fits, and likely revenue. Eventually one will ask you to call them first, and you can run one to three discussions to optimize price and terms.

Who should founders take money from?

I tell founders to take money from angels and VCs who can truly contribute. A company's path is up and down and up and down, and during the downs a CEO does not want to be alone or criticized. Good people around the table are positive and help the team out of the hole. We review what happened afterward.

Do rule-based approaches apply to controlling AI agents?

I would say the opposite applies. AI is probabilistic, so controlling AI agents requires a deterministic, rule-based approach. That may be controversial, but in the near future, rule-based control is key to making sure an agent does not deviate from what it is supposed to do. Four to six years from now it could be quite different.

What are you looking for in the next few quarters that you were not a year ago?

The world now better understands the difficulty enterprises have adopting AI. Almost all CEOs say they use AI, but it rarely improves the bottom line. Many agent pilots never reach production because teams lack security and compliance policies, cost control, good harnesses, and clean data. We are now investing in technology that controls agents or helps enterprises integrate AI.