Security
Attacker drained configs and root hashes from 996 Zyxel switches; CISA gives agencies until Thursday
CVE-2026-7273 was patched in June, but GreyNoise says exploitation began around August 17 across 48 countries.
Tuesday, September 22, 2026
Research
OpenAI says one internal model cleared 100 open math problems, and offers mathematicians an advisory seat
A new group hosted at the Institute for Advanced Study can comment on results and go public with its views, but cannot slow the research it is reviewing.
Security
Researcher publishes working exploit turning Meta's Muse assistant into a Mac backdoor
Patrick Wardle says an undocumented dictation setting can be rewritten by any local process, redirecting a user's spoken audio to an attacker's server.
Models
Grok 4.7 ships at Chinese-model prices and benchmarks well behind GPT-6 and Claude Fable 5.1
A larger base model and a longer reinforcement learning run leave the flagship mid-pack, and far back in agentic coding.
Business & Policy
UN science panel's first report on AI agents says there is 'no assurance humans will keep control'
Co-chair Yoshua Bengio says a real deployed system combined a misaligned goal, the means to pursue it and a permissive environment for the first time.
Business & Policy
US and China agree to an official AI dialogue days before the Trump-Xi summit
Bessent's proposed notification mechanism for national-security AI incidents is now attached to a standing channel, with chip export controls left out of the talks.
Business & Policy
Anthropic is building a biology lab where Claude directs robots through drug experiments
The San Francisco-area facility moves the company's life sciences work off simulation, with human oversight retained for safety and no clinical trials planned.
Business & Policy
Newsom signs seven bills making AI data centers pay for the grid upgrades they require
The California package targets energy and water use and is aimed at keeping utility costs off residents' bills.