Broader activity

Others hit in Hugging Face hack, companies endorse employee petition, Mythos cryptography breakthrough, and more

OpenAI's revised incident report says its unreleased model used publicly exposed credentials to reach accounts at four other services during the Hugging Face intrusion, including infrastructure provider Modal, whose misconfigured customer sandbox became a staging point; a parenthetical also admits similar account access during earlier evaluations, and METR argues independent researchers should be reviewing the evidence instead of the company itself (Robert)
After more than 1,000 frontier lab staff signed a petition for governance tools that would preserve the option of slowing AI development, Anthropic's Dario Amodei signed on and OpenAI endorsed the effort in a company statement; the wording still stops short of asking for an actual pause, which some think is overdue given recent events (Mitch)
The FCC banned imports of humanoid and quadruped robots along with power inverters on national security grounds, but existing carve-outs for previously approved Chinese inverter designs and doubts about how long the rules will last make a lasting effect on data centers or US robot makers unlikely (Mitch)
Anthropic reported that its Mythos Preview model found a way to break a candidate post-quantum encryption scheme in half the time of prior methods, using about 60 hours, a billion output tokens, and a few casually typed prompts; the two non-expert researchers needed close to a month to check work the model did in a week (Mitch)
The New York Times covered an Epoch AI projection that global AI computing capacity will grow tenfold by the end of 2028, with totals doubling roughly every nine months, US firms holding around 80 percent of capacity, and Google's Jeff Dean saying enough compute could fully automate the loop of AI improving AI (Mitch)