AI Cybersecurity Needs Openness, Not Secrecy

The cybersecurity world is buzzing about Mythos, the “frontier AI model” that can rapidly find and patch software vulnerabilities. Hugging Face just published a solid breakdown of what this means, and I’ve got some thoughts.

Let’s cut through the hype first. Mythos isn’t magic. It’s a large language model trained on code, but the real innovation is the system around it: lots of compute, tons of training data, scaffolding for vulnerability probing, speed, and some autonomy. That recipe—not any single model—is what makes it powerful.

And here’s the thing: others can build comparable systems. Smaller models wrapped in smart security engineering could produce similar results for less money. That’s great news for defenders. AI cybersecurity capability doesn’t scale smoothly with model size. The system matters way more than the model itself.

Why Openness Wins

As autonomous vulnerability-finding systems become common (and they will), open code and tooling level the playing field. Software security is now a speed race across four stages: detection, verification, coordination, and patch propagation. Open ecosystems distribute these across a community. Closed-source projects centralize everything inside one vendor—a single point of failure where only one organization can see and fix the code.

The distributed nature of open development is robust to that. Look at the Linux kernel security team, the Open Source Security Foundation, or Hugging Face’s own model security work. These communities move fast because many eyes see the code.

Proponents of closed systems often argue proprietary obscurity offers protection. That argument is getting weaker by the day. AI tools can now assist with reverse engineering stripped binaries. Most legacy firmware and embedded code is closed, binary-only, and no longer maintained. That’s a huge attack surface, and it’s becoming more legible as AI improves.

There’s another risk brewing inside closed codebases. Companies adopting AI coding tools under bad incentives (evaluating engineers by feature volume instead of code quality) can introduce more vulnerabilities than traditional development. Those vulnerabilities sit behind a single-organization firewall while AI-enabled attackers discover them from outside. That imbalance is exactly what open ecosystems avoid.

Underlying all this is capability asymmetry between attackers and defenders. Open models and tooling narrow that gap by giving defenders access to the same class of capabilities attackers can reach for—capabilities that would otherwise concentrate within a small number of well-resourced entities.

Semi-Autonomous Agents Hit the Sweet Spot

Based on the System Card, Mythos can operate with near-full autonomy. That’s something Hugging Face has advised against, and I agree. Loss of control is a real risk with fully autonomous systems.

Semi-autonomous agents strike a better balance. You prespecify what actions they can take and require human approval for certain steps. People stay in control, and the AI handles specific subtasks. This is doable with open code that organizations run privately, specifying allowable tools, skills, and access privileges.

With this setup, AI agents can be deployed defensively—finding vulnerabilities and assisting with patches—without the risks of full autonomy. It’s not as flashy as a fully autonomous system, but it’s where the real value lives.

The Bigger Picture

The Mythos announcement isn’t just about a new model. It’s about how we structure AI systems for cybersecurity. The recipe matters more than the model. Openness matters more than secrecy. Semi-autonomy matters more than full autonomy.

We’re entering a phase where AI can both attack and defend at machine speed. The choice isn’t between open and closed. It’s between distributed resilience and single points of failure. I know which side I’m on.