Claude helped researchers reach OpenAI's internal repos

The researchers trace two bugs: an SSO misconfiguration that exposed OpenAI accounts and internal repositories, and a heap overflow in libheif. The library is also used by Meta, Rails, Next.js, and ImageMagick, so the disclosure extends beyond one lab.

Unsealed filings expose internal warnings about AI training

TechCrunch

Unredacted filings show Microsoft and OpenAI staff describing publisher harm and the companies' data practices in stark terms. The exhibits add the defendants' own words to the copyright case instead of another round of outside argument.

The underlying court filing

U.S. District Court, Southern District of New York

CourtListener hosts the September 17 filing in the consolidated OpenAI copyright litigation. It is the primary docket entry behind the latest reporting.

The evaluator bench is small and contested

Axios

Axios maps the dispute over who can credibly audit frontier systems, including scrutiny of METR's ties to the AI safety community. METR says it doesn't take funding from frontier labs, and that Paul Christiano joined OpenAI's nonprofit board after its investigation concluded.

ISIL supporters are sharing AI safeguard bypasses

Al Jazeera

Al Jazeera obtained material showing ISIL supporters sharing ways to bypass safeguards in named AI systems. The report identifies a specific misuse community without detailing operational techniques.

A coalition goes looking for 100 gigawatts of grid capacity

TechCrunch

Google, Nvidia, Anthropic, and Emerald AI want to identify 100 gigawatts of grid capacity for new data centers. That target equals roughly a fifth of U.S. peak demand.

OpenAI and Anthropic pursue smaller data-center deals

CNBC

CNBC reports that both labs are considering smaller projects as they race to deploy capacity. These deals offer another route to power and compute while large campuses wait on grids and construction.

CXMT prepares to enter NAND flash

Reuters via Techmeme

Reuters reports that Chinese DRAM leader CXMT is preparing to expand into NAND flash amid memory shortages. The expansion would put it against Samsung, SK Hynix, and domestic rival YMTC in another part of the AI supply chain.

Solidigm considers a U.S. NAND factory

Reuters via Techmeme

SK Hynix subsidiary Solidigm is considering a U.S. NAND plant, while SK Hynix is reportedly discussing a separate Ohio project with Intel. Both talks address the same memory shortage with domestic capacity.

Stripe moves fraud detection ahead of the transaction

AI News & Strategy Daily

Stripe saw a sharp increase in command-line interface usage without a product change and attributed it to agents discovering the interface. Its response shifts detection toward abusive customers before they spend stolen tokens or consume costly inference.

Noam Brown says the 10,000-agent scaffold was a small part of the result

Dwarkesh Patel

Brown describes an experiment in which 10,000 agents ran for 88 hours. The run consumed 130 billion tokens. He attributes less than 10 percent of the success to the multi-agent system and says simple messaging primitives beat a rigid coordinator.

A cheap verifier catches most false passes

arXiv

A harness ablation matches fixed plans against shuffled policy text of equal length, separating useful guidance from prompt volume. Its read-only verifier rejects 61 percent of oracle-invalid retail episodes and withholds 17 percent of correct ones. The verifier costs less than one cent per episode.