Internal Files Surface Warnings Over Microsoft and OpenAI’s News Scraping

01New York Times and Other Media Disclose Microsoft and OpenAI Files: Employees Warned News Scraping Threatened Publishing’s Foundations

News organizations led by The New York Times have disclosed internal Microsoft and OpenAI documents that they say reveal warnings about using large volumes of news to train artificial-intelligence systems. The publishers accuse the companies of working together to take copyrighted journalism without permission, allegations Microsoft and OpenAI have contested through a fair-use defense.

The documents appeared in the publishers’ motion for summary judgment, which was unsealed Thursday after some of the material had previously been designated confidential. A summary-judgment motion asks a court to decide claims without a trial. The filing brings internal discussions into public view, but the quoted material and its legal significance are being presented as part of the news organizations’ case—not as a court finding that infringement occurred.

According to the publishers, Microsoft Director of Applied Science Brent Hecht repeatedly warned about the scale and legality of scraping news for AI training. They said Hecht described the practice as an unprecedented appropriation of labor and, in one document, possibly the largest such theft in history.

The plaintiffs also said Hecht questioned the companies’ central legal defense. Microsoft and OpenAI argue that training AI on news content qualifies as fair use, a doctrine that can permit some unauthorized uses of copyrighted material. But Hecht allegedly wrote that plans to scrape news widely made a “complete mockery” of that concept. The source does not establish how the court will interpret his remarks or how much weight it will give them.

The documents described a separate commercial danger for publishers. OpenAI’s Nick Turley, who leads ChatGPT, wrote in an internal message that commercial products trained on news could become substitutes for news providers, creating an “existential threat” for publishers, according to the filing.

One Microsoft document reportedly connected that substitution risk to the AI companies’ own dependence on publishers. It described a “doom loop” that could damage both model performance and the wider web: AI products weaken the economics of producing news, reducing the supply of material needed by large language models. The document said Microsoft’s LLM business had created an unusual situation in which its end product threatened the economic foundations of essential suppliers in its content supply chain.

02Anthropic Opens Life Sciences Verification Program, Giving Approved Teams Fewer Biology Research Blocks

Anthropic has opened applications for the beta version of its Life Sciences Verification Program, which gives vetted teams access to AI models with fewer restrictions on biology-related work. The company says its generally available Fable models can block legitimate tasks in drug discovery, research biology, clinical development and manufacturing because individual requests may resemble potentially harmful activity.

The program replaces some request-by-request blocking with institution-level verification and continuing monitoring. Anthropic has already admitted dozens of organizations through early access and is now accepting applications from academic laboratories, startups, pharmaceutical companies and other life sciences teams and institutions.

Applicants undergo reviews of their research credentials, security standards and ethical oversight. Approved organizations can seek either Standard Use or High-risk Use access, with permissions tied to the uses described in their applications.

Standard Use is intended for most biology research and development, including work in basic science, manufacturing, clinical development, quality assurance, regulatory affairs, investing and due diligence. It can cover an entire team, is renewed annually and provides Mythos 5.1, Opus 5 and Sonnet 5 with refined classifiers—automated safety filters—that are more permissive for scientific tasks. Anthropic says Standard Use will also apply to future models as they launch.

High-risk Use is an additional, project-specific grant for work still blocked under Standard Use. It removes all safeguards that block life sciences requests, but other protections, including cybersecurity classifiers, remain active. Each grant covers one research project and requires renewal every six months. High-risk access is currently available for Opus 5 and Sonnet 5; access for Mythos remains limited to a small group receiving additional vetting while Anthropic works with the US government on broader availability.

Both access levels work through Claude Science, Claude.ai, Claude Code and the API. Expansion to individual Pro and Max plans is planned, but Anthropic has not provided a date.

Instead of judging every biology prompt in isolation, Anthropic will continuously analyze activity across requests and sessions for patterns outside each organization’s declared scope. It can flag suspected unauthorized activity to institutional administrators, who must investigate and address incidents within previously agreed timeframes. The company presents this structure as protection against compromised accounts, malicious or coerced insiders, and unintended dangerous actions by AI agents, but has not disclosed approval rates or externally validated the program’s effectiveness.

03Nvidia Opens Native CUDA Kernel Development to Rust With Tile and SIMT Paths

Nvidia has announced CUDA Rust, a toolchain that lets developers write GPU kernels directly in Rust and compile them natively into PTX, Nvidia’s intermediate instruction format for CUDA GPUs. The move closes a persistent gap in an increasingly Rust-based AI systems stack: developers could already launch GPU kernels from Rust, but often had to write the kernels themselves in another language.

The company is involved in that shift through projects including its Rust-based Nova Linux driver and the Rust core of Nvidia Dynamo. Its new approach extends Rust into the GPU kernel layer rather than using it merely as a wrapper around code written elsewhere.

CUDA Rust offers two programming paths. Nvidia recommends starting with Tile, a model in which developers specify operations on a block, or “tile,” of data. The Tile IR compiler then decides how that work maps onto different GPU architectures, reducing architecture-specific choices in application source code.

SIMT, short for single instruction, multiple threads, provides finer control. It follows the familiar CUDA C++ and numba-cuda model: developers describe what one thread does and launch thousands of those threads. Nvidia positions SIMT for cases that require developers to manage threads and memory directly.

The SIMT implementation, called cuda-oxide, is a custom backend for the Rust compiler. It intercepts compilation and routes functions marked as GPU kernels through Rust’s Mid-level Intermediate Representation, the community-developed Pliron IR framework, and LLVM IR before producing PTX. Non-kernel code continues through the standard Rust backend. Host and device code can reside in one file and build with one command, without a separate kernel crate.

Nvidia demonstrated both paths with complete programs performing elementwise addition across 1,024 floating-point values. Both examples produced the same successful result, but that demonstration does not establish parity with the mature, enterprise-grade CUDA C++ or CUDA Python toolchains.

The current setup also has firm requirements: Linux, a GPU with compute capability 8.0 or later, CUDA Toolkit 12.x or newer, clang and its libclang headers, and a pinned nightly Rust toolchain. Nvidia says it plans to support cross-language interoperability and continue maturing CUDA Rust through 2027 and beyond, but has not specified a stable-release timetable or the scope of production support.