Saturday, September 12, 2026
A researcher timeline puts more than 2,000 malicious uploads on May 11 and 12, forcing maintainers to freeze sign-ups for four days, two months before the Hugging Face hack.
Users in northern Yemen returned to Claude to ask why their guided rocket failed, while a French-speaking ShinyHunters member ran secret-scanning across ten AWS workers.
CVE-2026-86793 chains Python builtins to escape the inference framework's SafeUnpickler; any deployment without an API key configured is exposed.
Terence Tao and fellow signatories say lab goals are severely misaligned with the field, as OpenAI withdraws sponsorship of a Caltech math event.
With a morality briefing, five of six reasoning models kept kill rates under 6 percent; without it, every model exceeded 84 percent.
The deep learning pioneer says better goal optimization brings better deception, rule-gaming and coordination, and renews his call for independent safety reviews before deployment.
The new command runs realistic prompts with and without a plugin loaded, grades the output with six grader types, and can gate skills in CI.
Sam Altman told staff the company could slow its pace alongside other labs, and chief scientist Jakub Pachocki has called for a coordinated pause until shared safety standards exist.