security
Your LLM Security Moat (Or Someone Else's)
Everyone’s deploying AI agents. Most will fail.
Not because the models aren’t smart enough—they are. METR’s latest benchmark shows agents can now handle nearly 5 hours of human-equivalent work. By year-end, we’re looking at 40-hour autonomous work weeks. The capability curve isn’t exponential anymore. It’s super-exponential.
The failure won’t be intelligence. It’ll be trust.
The Moat Is Not the Model
Nate Jones has been hammering this point: your competitive advantage in 2026 isn’t having access to Claude or GPT. Everyone has that. The models are commoditized.
Your moat is your domain-specific infrastructure—the schemas, verification loops, and guardrails you build around those models. The stuff that knows your business, your data, your edge cases.
Generic AI security is like generic ITSM: designed for everyone, fits no one. You might as well pull a yellowing process chart off the wall and scan it into your LLM. You’d learn more from the bright patch of blank wall it left behind.
Build Your Own
Today I’m opening up Injectionator Labs.
Two things you can now do:
-
Develop your own injectionator — Your prompts. Your attack patterns. Your domain knowledge baked in. The Labs environment lets you build security that actually understands what you’re protecting.
-
Deploy it yourself — Cloudflare Workers or containers. Your infrastructure, your control, your data never leaving your perimeter.
You’re not buying a product. You’re building infrastructure. The moat becomes yours.
The Human Throttle Problem
Here’s the uncomfortable bit: as agents get faster, humans become the bottleneck. Nate calls this the “Human Throttle Problem.” We can’t review at machine speed. We can’t approve at machine speed.
The only way through is building systems you actually trust to run autonomously. That means security that’s specific enough to catch real threats and permissive enough to let legitimate work flow.
Generic guardrails will either block everything useful or miss everything dangerous. Domain-specific security—built on your data, tuned to your context—is what lets you take the throttle off.
I wrote more about this in The M in MVP Has Changed—how Jevons Paradox means cheaper AI doesn’t mean less AI work, it means radically more. And more AI work means more attack surface, unless you’re building security that scales with it.
The Pioneer Moment
This is the frontier. The settlers and town planners come later. Right now, you’re either building your own infrastructure or you’re training someone else’s.
Your security moat, or theirs.