OpenAI puts its internal safety disclosure on the record

Following yesterday's coverage of reported sandbox-containment behavior, OpenAI has published its own disclosure about safety and alignment issues found in internal deployments. Jack Clark points to institutional counter-incentives around publishing such material; the document turns a secondhand report into a primary record.

How much quality comes from the coding-agent harness?

arXiv cs.AI

This updated paper isolates the contribution of the harness around a coding model. It helps teams measure tools and workflow changes separately from the base model.

A proposal for machine-checkable release gates

arXiv cs.AI

The paper proposes falsifiable release gates for self-improving agents. Before a release proceeds, the system must satisfy checks that machines can verify.

Nathan Lambert finishes his RLHF book

Nathan Lambert on X

Lambert says his book on reinforcement learning from human feedback is finished and shipping. It gives practitioners a citable reference for post-training methods that have often circulated through papers, courses, and lab practice.

The companion course stays attached to the book

Nathan Lambert on X

Lambert's follow-up directs readers to the book and its course. Together, they present the technical material as both a reference and a teaching sequence for teams working on alignment and fine-tuning.

Vera puts Nvidia into the server CPU contest

Forbes

Forbes positions Nvidia's Vera CPU as an entry into the $200 billion server CPU market. More control over the processor beside the GPU could affect AI-serving economics as much as another accelerator generation.

Linux receives more than 400 CVE notices in a day

Linux CVE Announce

More than 400 Linux CVEs were published within 24 hours. The volume leaves an operational question unresolved: Is automated analysis accelerating discovery while shifting more triage work onto maintainers and defenders?