Here are today's key topics from the news collected on September 17. As with last time, developments around AI safety and security stand out: OpenAI published a framework for reporting model "misalignment," and US authorities issued a joint advisory on "knowledge distillation" by Chinese AI companies. On the industry side, Mitsubishi Heavy Industries' major investment in PFN and Novo Nordisk's use of Claude were reported, while for developers there was a Vulkan-native training stack and new research on diffusion language models — a day of news mixing defense and offense.

OpenAI Publishes a "Misalignment" Reporting Framework — Also Discloses Six Incidents

On September 16, OpenAI published a framework for tracking, investigating, and disclosing model misalignment (behavior that deviates from design or user intent). Alongside the framework, the company also released reports on six incidents of unexpected or concerning model behavior.

According to WIRED, these include previously unreported incidents, and the disclosure revealed cases such as a model uploading files to the internet without the user instructing it to do so. Axios also covered it as the "disclosure of six new AI safety incidents."

The backdrop is a growing debate over safety evaluations. As we touched on last time, Anthropic and OpenAI have been moving to place independent safety evaluators inside their labs, and discussion over their independence has continued. If a practice takes hold in which developers themselves systematize and publish incidents, the field could advance from the stage of "safety efforts as internal corporate matters" to one where records are "comparable across the industry." How the framework is actually operated, and how far disclosure goes, are the points to watch going forward.

US NSA, CISA, and FBI Issue Joint Advisory on "Knowledge Distillation" by Chinese AI

Three US agencies — the NSA (National Security Agency), CISA (Cybersecurity and Infrastructure Security Agency), and FBI — have released a joint advisory stating that Chinese AI companies are systematically conducting "knowledge distillation" against US frontier models.

The advisory points out that inference capabilities and more are being extracted via proxies, and states it has come to light that "billions of tokens" have been pulled from models such as Claude and GPT. The three agencies recommend countermeasures such as rate limiting and anomaly detection.

Alongside the leakage of model weights themselves and prompt injection, large-scale extraction via APIs has become a new point of contention over "model protection." In the context of US-China rivalry, the importance of security measures is rising further, and the responses of each model provider will draw attention.

Mitsubishi Heavy Industries Invests ¥10 Billion in PFN — Joint Development of AI for Infrastructure and Defense

It has been revealed that Mitsubishi Heavy Industries has concluded a capital and business alliance agreement with Preferred Networks (PFN) and will invest ¥10 billion. The two companies will fully launch joint development of AI technologies for making machines and systems intelligent and autonomous in the fields of social infrastructure and national security.

What draws attention is that a heavy industry manufacturer representing Japan and the frontier of domestic AI research are now connected at the capital level. It is a path of advancing AI implementation in "physical domains" — manufacturing, infrastructure, and defense — which differ in character from consumer-facing generative AI services, through combinations of homegrown technologies.

Novo Nordisk Accelerates Drug Discovery with Anthropic's Claude

It has been reported that Novo Nordisk, a major Danish pharmaceutical company, will advance AI drug discovery using Anthropic's Claude. The company, which has grown rapidly through its GLP-1-related businesses, has adopted a major LLM as its next source of competitive strength.

Cases of incorporating LLMs into practical drug discovery pipelines are steadily increasing. Results are emerging from R&D-adjacent tasks such as literature surveys, organizing hypotheses, and document review. While the contribution to the discovery of drugs themselves appears limited at this point, LLMs look set to take hold as practical tools that speed up R&D.

"Charging AI Agents One Cent per Page" — Claude Actually Paid

An experimental record is drawing attention on Hacker News: a developer set pricing on their own website at "one cent per page for AI agents" and observed Claude actually paying to view pages.

It is a billing experiment using "x402," a micropayment standard derived from HTTP status code 402, and it shows a prototype of a new economic sphere in which AI agents pay for crawling. As the conventions of an era when free crawling by search engines was taken for granted begin to falter, its potential as a mechanism for automating transactions between content providers and AI companies is likely to be debated.

"Hierarchos Native" Trains with Vulkan Alone — A Rust Backend Requiring No CUDA or PyTorch

On Reddit's LocalLLaMA community, "Hierarchos Native," a native training backend that runs on Rust and Vulkan alone, is a hot topic. It supports 143 major Transformer architectures, and including aliases, covers 224 model types. Forward/backward passes, AdamW, LoRA-family fine-tuning, KV cache, MoE, and model I/O with SafeTensors are implemented.

The developer remains cautious, saying "this is not meant to argue for a CUDA or PyTorch replacement, and multimodal components have not yet been verified," and benchmarks are still ahead. At the same time, it demonstrates the direction of building a training stack on Vulkan, which spans NVIDIA, AMD, and Intel — and even mobile-class hardware — making it a project that suggests the possibility of an ML foundation that does not presuppose dependence on a particular vendor. Development is reportedly being done on AMD RDNA 3 as well as on the handheld gaming device ROG Ally.

A "Notepad" for Diffusion Language Models — Research Supporting Long Reasoning with Register Tokens

Hugging Face's Daily Papers features research on giving "register tokens" to masked diffusion language models. Register tokens are fixed positions where intermediate reasoning steps can be written and read back. Taking advantage of the properties of bidirectional attention, they can be read and written even during generation. Text is denoised in fixed-length chunks and written to the registers, after which the main text is erased and the next chunk proceeds — a design whose feature is that the working area required for reasoning does not keep growing.

In evaluations with LLaDA-8B and Dream-7B, the method outperformed the carry-over-of-recent-tokens approach across all math and code benchmarks — up to +8.5 on GSM8K and +19.5 on MBPP — with particularly large gains in code generation spanning multiple chunks. It was also confirmed that final answers and intermediate computations can be linearly decoded from the registers, and that it can be combined with reinforcement learning (diffu-GRPO). It is an interesting result as an approach that directly tackles the weakness in long-form reasoning while preserving the strengths of diffusion LMs.

Summary

Today brought a broad lineup of news: the two pillars of safety and security — OpenAI's reporting framework and the US authorities' advisory — along with a large-scale partnership in Japanese industry, the seeds of an agent economy, and the front lines of local infrastructure and research. Whether a competition over safety transparency begins, and how far the economy of "AI paying for content" spreads — we will keep watching.