---
title: "Field notes, November 9, 2026 — Sekura"
url: https://sekura.ai/notes/2026-11-09/
description: "2 notes from Tikka Nagi on application security and AI: Paul Graham is not writing about startups; An LLM can write a brilliant zero-day"
updated: 2026-11-09
generator: sekura build, scripts/build-markdown.mjs
---
Field notes

# November 9, 2026

Short takes on application security and AI, also posted on X and LinkedIn.

[All field notes →](https://sekura.ai/notes/)

![Sekura — owl-in-shield mark](https://sekura.ai/branding/sekura-owl-v3.png)

November 9, 2026 · Also posted on LinkedIn · [#](https://sekura.ai/notes/2026-11-09/#j-102)

Paul Graham is not writing about startups. He is writing about the physics of leverage.

Most founders obsess over features. Graham obsesses over distribution. He understands that code without distribution is just a hobby. He built Y Combinator not as a fund, but as a compiler for founder potential.

This matters because AI is lowering the cost of creation. The bottleneck is no longer writing code. It is finding a market that actually pays. Graham’s essays remain the most efficient map through that noise. He cuts through the VC theater to the one truth that survives: solve a real problem for a real person.

Stop reading pitch decks. Start reading his essays. They are free. They are brutal. They are the only cheat code that works in every cycle.

---

November 9, 2026 · Also posted on LinkedIn · [#](https://sekura.ai/notes/2026-11-09/#j-83)

An LLM can write a brilliant zero-day. It can also write a hallucinated one.  
The problem is that the model cannot tell the difference.  
It predicts tokens. It does not predict crashes.  
This is why autonomous pentesting fails when it relies on LLM self-certification.  
The model proposes the attack vector. A deterministic Rust engine attempts the exploit in a sandbox.  
If the target does not break, the finding does not exist.  
We call this VerdictGate.  
It separates AI creativity from AI verification.  
Creativity is infinite. Verification is binary.  
Stop asking the model if it worked.  
Make the code prove it.

[All field notes](https://sekura.ai/notes/)
