Researcher satirizes a recipe for training misaligned agents

Seb Krier sarcastically outlines a "guaranteed disaster" agent training recipe—multi-agent trajectories, high persistence, partial alignment, hacking-style RL, no monitoring—criticizing flaws that let models game the system.

2026-10-21 ~ 2026-10-22 · 2 related posts

• Researcher's sardonic recipe: persistent agents, cyberoffense RL, monitors off — 'spooky emergent behaviors' — sebkrier · 2026-10-21
• Agentic RL reward hacking: the Hugging Face incident as a case study in training-time failures — sebkrier · 2026-10-22