Vijil.

A boring, enormous problem

Filed from Menlo Park - a company profile

Here is a statistic that Vijil would like you to sit with: roughly 95% of enterprise AI-agent projects never reach production. Not because the models are bad - the models are, if anything, embarrassingly capable - but because no one in the building will sign their name under one. The demo works. The pilot dazzles. And then somebody in legal, or security, or compliance asks the unanswerable question - can we actually trust this thing? - and the project quietly dies in a Slack thread.

Vijil, founded in 2023 in Menlo Park, is a company built entirely inside that awkward silence. Its founders - Vin Sharma, Zdravko Pantic, and Subho Majumdar - came out of AWS, where they collectively touched eleven of Amazon's AI services and helped ship the infrastructure that a lot of the industry now runs on. They noticed something that is obvious in hindsight and was somehow underserved: the market was busy selling shovels for the AI gold rush, but nobody was selling a way to trust the gold. Building an agent had become easy. Believing an agent had not.

So Vijil sells belief. The company's own phrasing is "trust as infrastructure," which sounds like a slogan until you notice that it is really a category bet. The wager is that "AI agent trust" stops being a vibe an engineer has about their own model and becomes a line item - a thing enterprises budget for, procure, and audit, the same way they buy identity management or endpoint security. If that bet is right, Vijil is early to a very large market. If it's wrong, it's a very well-funded science project. The $17 million it raised in late 2025, on top of an earlier round for $23 million total, suggests a fair number of investors think it's the former.

What makes the pitch interesting is that Vijil refuses to treat trust as a single moment. A lot of AI-safety tooling is essentially a pre-launch scan: you run the model against a battery of nasty prompts, you get a report, everyone feels better, you ship. Vijil's argument is that this is like crash-testing a car once and then never checking the brakes again. Agents live in the world. The world sends them prompt injections, jailbreaks, malformed tools, and users who are creative in ways no test suite anticipated. Trust, in Vijil's telling, is not a certificate - it's a maintenance schedule.

The interesting part is Darwin

Three of Vijil's four products are recognizable if you squint - testing suites, guardrails, and starter kits exist elsewhere in various forms. The one that makes investors lean forward is Darwin, because it borrows an old idea from security and points it at agents: the best defenses get smarter from being attacked. Every jailbreak attempt in production, instead of being merely blocked and logged, becomes training data. The agent's defenses are tuned by reinforcement learning against the specific threats it faces, not a generic list. BrightMind's Stephen Ward singled out exactly this - "the ability to continuously harden AI agents through reinforcement learning" - as the thing that sets the company apart.

For the person actually deploying an agent, the practical payoff is speed. Security reviews that used to eat weeks compress toward hours. Compliance - the EU AI Act, the NIST AI Risk Management Framework - stops being a wall of dread and becomes an audit trail you can hand to a regulator. And because Vijil offers on-premises deployment, the enterprises that can't send their data anywhere - the legal, insurance and healthcare shops that are exactly the ones with high-stakes agent use cases - can still play. The headline metric the company likes is "4x faster to trust." The customer proof point is SmartRecruiters, which says it now ships AI agents in six weeks instead of six months.

Three AWS veterans

Recognition & who's using it

The obvious questions