OpenAI holds its biggest training runs + the Cerebras CS-4

Eleven days after flagging that its unreleased Astra model might have crossed the "Critical" cyber tier, OpenAI has published what it built in response — and what it is still not running. Training now carries chain-of-thought monitoring aimed at establishing "what the model's actual goals are," automated alerts to safety staff within 30 minutes, and an automatic training halt if those teams cannot clear an alert in that window. The safeguards cost roughly 20 percent extra compute. Lower-risk work resumed after a two-week stop; the largest planned frontier runs are still on hold.

Cerebras announced its CS-4 rack system, claiming more than 1,000 tokens per second on models above 10 trillion parameters and up to 30 times the speed of production graphics-processor systems. The headline part is what the chip is not: the WSE-3 Turbo is the same 900,000-core, 5-nanometer wafer as the two-year-old WSE-3, clocked from 1.4 gigahertz to 2.8 gigahertz rather than re-fabricated. OpenAI is using it for the Ultrafast tier, and AMD is pairing its graphics processors for prefill with Cerebras for token generation. Shipments start this quarter.

The transformer-specialized chip startup Etched raised $700 million, led by Jane Street, at a $21 billion valuation — roughly double the $10.3 billion it carried in July. Unusually for the quant fund, Jane Street is also a customer: it has installed Etched's first shipped cluster in its own data center and says it is "pleased with the early results." Etched now sells a two-part system, a prefill chip plus a cluster-scale memory design, rather than a single accelerator.

Cursor, now owned by SpaceX, has started hosting code. Origin is in early beta on every paid plan and does repositories, browsing, and pull requests, with comments syncing both ways to GitHub within seconds so teams do not have to migrate. The timing is pointed: GitHub logged 257 outages over the past year, and one on the day that Origin launched pushed error rates near 20 percent worldwide.

Varonis Threat Labs kept asking Copilot why auto-execution would not work, and Copilot answered — naming the disabled parameters and one undocumented one, autorun=1. Appending that to a crafted query string produced a zero-click prompt injection: a victim clicks a link, Copilot loads in their authenticated session, and the attacker's instructions run with no prompt and no visible sign. Reachable data included mail, calendar, chat history, connected drives and Copilot's stored memory. Microsoft planned a patch and a formal vulnerability identifier for Tuesday.

Modular put the entire Mojo language — compiler and tooling — under the unrestricted Apache 2.0 license, a week after Mojo reached version 1.0. It is the first Mojo release since Qualcomm completed its acquisition of Modular on July 29, and Modular says the platform will keep optimizing for hardware that competes directly with Qualcomm's own. The same keynote added support for Qualcomm's Cloud AI 100 and Dragonfly accelerators.

ChatGPT for Teens began rolling out globally, and teenagers do not opt in — anyone who states an age of 13 to 17, or whom OpenAI's age-prediction system estimates is under 18, is moved over automatically. The model will not use romantic language, imply feelings, call itself a friend, or encourage emotional dependence, and a Study Hours setting can force Study Mode on a schedule. OpenAI has not published how accurate its age prediction is, and Common Sense Media maintains that social AI companions are not safe for children at all.