depthfirst.

Here is a fact about software security that is both obvious and slightly depressing: the amount of code in the world is going up, the speed at which it is written is going up, and - thanks to AI writing a lot of that code - both curves are getting steeper. Meanwhile the number of humans who can carefully read code for the one line where a bug becomes a breach is going up much more slowly, if at all. This is a supply-and-demand problem, and depthfirst's entire pitch is that you cannot solve it by hiring. You solve it by pointing machines at the problem, because the machines are also what created it.

depthfirst is a San Francisco company, founded in 2024, that calls itself an "applied AI lab." That phrase is worth pausing on, because it is doing two jobs at once. "Applied" means: we ship a product you can buy, today, that plugs into your developers' pull requests. "AI lab" means: we also train our own models, and we would like you to think of us as a research shop, in the way you think of the frontier labs. Most security companies are one or the other. depthfirst insists on being both, and it has raised $120 million to prove the combination isn't just a slide.

The product is organized around an idea the company brands as "General Security Intelligence" - a deliberate, slightly cheeky play on "artificial general intelligence." The claim is not that depthfirst has built a magic brain. The claim is narrower and more useful: an AI agent that reads your code, your business logic, and your infrastructure together can find vulnerabilities that a scanner looking at any one layer in isolation will miss. A traditional static analysis tool sees a function. depthfirst's agents are supposed to see the function, understand what the application does, and reason about whether an attacker could actually reach that line and do something bad with it.

If you have ever run a security scanner, you know the real enemy is not missed bugs. It is false positives - the thousand red alerts that turn out to be nothing, which train your developers to ignore all of them, including the one that mattered. depthfirst's headline numbers are aimed squarely at this. The company says its agents surface roughly 8x more true-positive vulnerabilities than traditional static analysis while cutting false positives by about 85%. Those are the company's own figures, and any buyer should test them, but the direction is the interesting part: more signal, less noise.

The number that actually reveals the strategy, though, is this one: depthfirst says 80% of its fix recommendations get accepted and merged by developers. This is a subtle but important choice of metric. Plenty of tools can tell you something is wrong. Far fewer can write the fix, in your codebase, in a way an engineer looks at and thinks "yes, ship it." The fix-accept rate is a trust metric. It is the difference between a tool that generates work and a tool that removes it.

How do you build agents that developers trust? depthfirst's answer, and the reason it needed the second $80 million, is that you train your own models. The company has built a specialized security model it calls dfs-mini1, which it first pointed at cryptocurrency smart contracts - a domain where a single overlooked bug can drain a wallet in seconds and where adversaries are relentless. depthfirst says the model runs at 10-30x lower cost than frontier models while outperforming them on OpenAI's EVMBench. The bet embedded here is the whole company's thesis: for the highest-stakes, most repetitive slice of a real workflow, a small model trained on exactly that task beats a giant model trained on everything.

The founding team is the kind that makes investors return phone calls quickly. CEO Qasim Mithani built developer and security products in the early days of AWS, then became an early leader at Databricks, where he built the infrastructure and enterprise platform from the ground up. CTO Andrea Michi was a Research Engineering Lead at Google DeepMind, where he worked on reinforcement learning for Gemini and AlphaDev and published in Nature. Co-founder Daniele Perito co-founded the wholesale marketplace Faire, led security and fraud at Cash App in its early days, and did postdoctoral security research at UC Berkeley. That is a security operator, a frontier-model researcher, and a repeat founder in one room, which is roughly the exact resume you would design if you were trying to build "an AI lab, but for security."

Investors agreed, and did so at a speed that is itself a small story. Accel led a $40 million Series A that depthfirst announced in January 2026 as it emerged from stealth, with backers including Jeff Dean, Kirsten Green, and Logan Kilpatrick. Then - less than 90 days later - Meritech Capital led an $80 million Series B, with Forerunner Ventures and The House Fund joining, at a reported $580 million valuation. Two rounds in a single quarter, before the company turned two. In venture terms this is a market voting with unusual conviction that AI-native security is a category, not a feature.

What can you actually do with depthfirst? If you run an engineering org, you point it at your repositories and it reviews pull requests in real time, flags reachable supply-chain risks (not just every dependency that exists, but the ones an attacker can actually touch), catches leaked secrets before they ship, and runs agentic pentests against your running code. When it finds something, it doesn't just file a ticket - it proposes the patch. The company says customers already include ClickUp, Supabase, Lovable, incident.io, Moveworks, Persona, AngelList, and Corelight; Persona has said it doubled its code-security coverage using the platform.

The honest caveat, which depthfirst would probably agree with, is that the numbers here are early and largely self-reported, and "AI agents that autonomously secure your software" is a sentence that has to survive contact with real, weird, adversarial production systems. Agentic pentesting still wants a human in the loop. Model claims want independent benchmarks. But the underlying observation - that software is now written faster than humans can review it, and that the only thing fast enough to keep up is more software - is not really in dispute. depthfirst's wager is that it can be the software that keeps up. It has $120 million and a very good founding team to find out.