Posts

What Actually Breaks When LLMs Write Code?

June 11, 2026 · Sumon Biswas

Our new preprint studies 547 real-world safety failures of agentic code assistants — not jailbreaks, but ordinary tasks gone wrong. A reflection on why operational safety deserves as much attention as adversarial safety.

Teaching LLMs to Plan Before They Act

June 10, 2026 · Sumon Biswas

Reflections on our ICML 2026 paper, Plan Then Action: why token-by-token chain-of-thought tends to wander, and what changes when a model is trained to commit to a high-level plan before it starts reasoning.