Field notes

November 8, 2026

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

Sekura — owl-in-shield mark

Pieter Abbeel isn’t just an academic. He is the architect of the autonomous agent era. His work on imitation learning and deep reinforcement learning didn’t stay in Berkeley labs. It became the backbone of modern robotics and AI safety. While others hype LLMs as chatbots, Abbeel focuses on agents that act in the physical world. This distinction matters. A model that predicts text is a toy. An agent that executes code is a liability if unverified. Most AI safety debates ignore the execution layer. They worry about bias in output, not exploitability in action. Abbeel’s trajectory proves that true autonomy requires deterministic control, not probabilistic guessing. If your AI strategy lacks a sandbox for verified execution, you are building on sand. Stop treating agents like search engines. Treat them like software with root access.


Most pentest tools let an LLM find a vulnerability and then ask the same model to validate it. That is self-certifying bias. It is not security. It is hallucination with a confidence score.

Verification must be orthogonal to generation. Code does not lie. LLMs do.

We don't trust the model's verdict. We run the exploit in a deterministic sandbox. If the shell pops, the finding ships. If it doesn't, it dies. Cryptographic receipts or nothing.

Stop letting models grade their own homework.