AI Security Over the Long Horizon | Aaron Brown | Misaligned 2026

Aaron Brown at Misaligned 2026 with the most hands-on talk on the program: how to actually train a model toward long-horizon security work, on hardware a small team already owns, with the code in the open. The claim underneath the demo is that the bottleneck is not compute and it is not the base model. It is the verifier. Most people skip straight to reinforcement learning with a reward function they borrowed, and the model dutifully learns to satisfy the wrong thing. Aaron's instruction is unglamorous and correct: build your own verifiers, and score by hand first. If you cannot tell a good trajectory from a bad one yourself, no amount of training will do it for you. The environment he built for it, Open Trajectory Gym, is open and the repository is up. This is a talk you can go and reproduce, which is a rarer thing on a security stage than it should be. Speaker: Aaron Brown, Founder & CEO, Stealth startup Misaligned is an invitation-only conference on AI and offensive security, held at the Mob Museum in Las Vegas during Black Hat USA week. A small room, a single track, and talks from the people doing the work: no sales pitches, no product demos. The program is professionally filmed so the room can stay small while the audience does not. More: https://misalignedcon.org #AISecurity #ReinforcementLearning #OffensiveSecurity #MLSecurity #BlackHat2026