Build for Where AI Is Heading

I use FFmpeg to preprocess dashcam video for TARA, a transport analysis tool I built. FFmpeg is one of the most widely used pieces of software in the world. It handles video processing in millions of applications.

On 8 April 2026, Anthropic announced that an unreleased AI model called Claude Mythos Preview had found a 16-year-old security flaw in FFmpeg. Automated testing tools had run past the bug five million times without catching it. I don’t know the details of the flaw or whether it affected my usage. But software I depend on had a hidden problem for 16 years and nobody knew.

That is one example of what Mythos-level intelligence can do. The security findings are what made the headlines. What I keep thinking about is what this level of capability means for building.

What Anthropic announced

Anthropic built a model that is significantly better at reading, reasoning about, and working with code than anything available today. On standard tests that measure how well AI can understand and fix software, Mythos scores 93.9% where their current best public model scores 80.8%. On cybersecurity-specific tests, 83.1% compared to 66.6%.

They used the model to find thousands of previously unknown security flaws in every major operating system and browser, then formed a coalition of twelve organisations to fix them: AWS, Apple, Google, Microsoft, CrowdStrike, Cisco, JPMorganChase, NVIDIA, Palo Alto Networks, Broadcom, and the Linux Foundation. Anthropic committed $100M in credits and $4M to open-source security. The project is called Glasswing.

Anthropic chose not to release Mythos publicly. The model is too capable to put out without safeguards that do not yet exist. They plan to develop those safeguards on an upcoming model first, then apply them to Mythos-class capabilities.

What caught my attention

The security headlines are important. But the thing that stayed with me is more basic. This model can read a complex piece of software, understand how all its parts fit together, find a flaw that five million automated tests missed, and figure out how to use that flaw. Autonomously. No human steering.

That is a level of reasoning about code that changes what is possible for people who build things.

TARA uses AI today to analyse dashcam video and generate transport appraisals. The AI handles the video. I handle the economic logic, the report structure, the domain knowledge. With Mythos-level intelligence, that division of labour shifts. A model that can reason about code at this depth could handle more of the economic analysis itself. It could audit its own outputs. It could spot flaws in my logic the way it spotted flaws in FFmpeg.

I am not speculating about a distant future. The benchmarks show this capability exists today, behind closed doors. If history is any guide, it will be widely available within a year. Almost certainly within two.

What this means if you build with AI

The ceiling on what one person can build with AI keeps rising. Six months ago I could not have built TARA alone. Today I can. Six months from now the scope of what a single builder can tackle will expand again. A year from now, again.

Think about what you build today and where AI helps. Now imagine the AI component getting meaningfully better at understanding your domain, reasoning about complex problems, and working through multi-step tasks without hand-holding. What becomes possible?

A finance professional building dashboards gets a model that can audit its own calculations. An engineer building analysis tools gets a model that understands the engineering standards, not just the code. A civic worker building permit workflows gets a model that can reason about the regulations themselves.

The gap between “I have an idea for a tool” and “I have a working tool” is shrinking fast. Mythos-level intelligence means it shrinks further. The people who benefit most will be the ones who understand their domain deeply and can direct AI toward real problems. The technical barrier keeps falling. The domain knowledge becomes the advantage.

Build for where it is heading

Security is one obvious implication of all this. Every tool you build will exist in a world where AI can find its flaws faster than you can fix them. That is worth taking seriously from day one.

But the bigger point is about pace. The assumptions you make today about what AI can and cannot do have a shorter shelf life than you think. How your tools are designed, what you plan to build next, your sense of what is feasible for one person or a small team. All of it needs to account for capability jumps that are getting larger and more frequent.

I build a transport analysis tool that depends on software with a 16-year-old flaw nobody knew about until an unreleased AI model found it. A year from now, that model’s capabilities will be routine. The question for every builder is: what will you build when they are?

How this was written: Ideas and key arguments are the author’s. AI assisted with research and drafting.