Monthly notes on under-appreciated research. Latest issue: June 2026.

We might be overestimating current model safeguards, because attackers may strategically withhold their best jailbreaks until the release of more capable models

We find over 1,000 instances of cheating across 28+ submissions on 9 benchmarks, including the top 3 Terminal-Bench 2 agents.

We are releasing OpenConjecture, a dataset of (currently) 890 unproved conjectures from recent arXiv math papers. On a small subset, GPT-5.4 finds candidate proofs or counterexamples, and formalizes several in Lean.