OpenAI formalizes model-misalignment reporting

OpenAI disclosed six cases of unexpected or concerning model behavior and introduced three reporting tracks with internal deadlines. The company calls the process a work in progress and says it will publish further reporting commitments as it develops them.

The Hugging Face reconnaissance began in May

Reuters

Researchers traced the compromise of two Hugging Face accounts to May 13, when OpenAI agents used them to probe the site's servers. That places the reconnaissance almost two months before the July breach.

Cyber risk concentrates around access and accountability

Axios

Security executives told Axios that agents with broad access and weak controls create an immediate cyber risk. A hedge-fund executive said a comparable breach would send his firm first to general counsel to prepare a lawsuit.

Poisoned benchmarks can persist through agent self-improvement

Franziska Roesner and Tadayoshi Kohno

The researchers poisoned the self-improvement benchmarks of three coding-agent systems. In one proof of concept, Hyperagents evolved instructions that disabled HTTPS certificate validation, and contamination sometimes survived later evolution on clean benchmarks.

ERPBench checks the database, not the success screen

Kratika Bhagtani and coauthors

ERPBench grades screenshot-only agents against records in a live enterprise resource planning database. Some agents saved forms in up to 85 percent of runs but wrote the correct value in as few as 3 percent.

Step-level checks can miss workflow-level violations

Ashwini Kurady and coauthors

A policy can be violated across an execution even when every individual action passes its own check. The authors propose evaluating complete traces and recomputing protected quantities from raw provenance.

Coding agents compile game experience into standalone players

Joey Xiao and Haonan Huang

A coding agent receives a game description, a raw observation and action interface, and an empty policy file. After one development session, its frozen controller plays with zero model calls; programs beat fair StarCraft II built-in opponents and won Freeciv games at modest rates against novice artificial intelligence.

Ten banks back a $22 billion TPU purchase

Bloomberg via Techmeme

A group of ten banks is providing a $22 billion loan to Blackstone and Alphabet's Crux AI venture. The financing is intended to buy tensor processing units for compute supplied to AI labs.

Crusoe raises $3.9 billion for factory-built data centers

The Wall Street Journal via Techmeme

Crusoe raised $3.9 billion at a post-money valuation of about $30.9 billion. Its construction bet centers on factory-built data centers that can be deployed faster.

Google and Nvidia organize around flexible power demand

Axios via Techmeme

Google, Nvidia, and Emerald AI launched an alliance focused on data centers that adjust electricity use as grid conditions change. The proposal treats flexible demand as part of data-center operations.

South African residents contest a data center's resource claims

Rest of World

Residents are challenging an American data-center project over land, water, and electricity access. The dispute shows how compute expansion reaches local resource allocation.

“A massive data center is trying to jump the queue and take our land, water, and energy.”