AI research

On October 29, Randall Balestriero and Yann LeCun posted a paper to arXiv with a title that dares you to skim it: Proving and Improving the Platonic

One shared ChatGPT link ran attacker code in your session. The sandbox held — the hole was a postMessage bridge, still unadvised upstream.

Anthropic official Claude Code extension shipped a 66-line SSH-backdoor loader through the Cursor marketplace for 11 hours. The primary sources, checked.

A frozen-model harness benchmark measures two things at once. arXiv 2610.37255 splits them exactly: return = ceiling minus mismatch.

A six-agent lab passed all ten accuracy criteria, then refused the run anyway - the controls failed on room-light drift. What precommitted criteria buy you.