Three Mile Island Nuclear Plant to Restart Under Microsoft AI Power Deal; Multi-Stage Attack Breaks Diffusion LM Watermarks; Colopl Releases Free Creator Protection Tool

Three Mile Island Nuclear Plant Moves Toward Restart Under Microsoft's AI Power Deal

The Three Mile Island nuclear power plant, famous for its 1979 accident, has begun moving toward a restart backed by a power supply contract for Microsoft's AI data centers. According to Bloomberg, the restart plan has entered a concrete phase, with the surge in power consumption driven by AI demand giving the plant's revival a push.

The power required to train and run AI models grows year by year, and major technology companies have begun to position nuclear energy as a "clean baseload power source." Microsoft had already announced a power purchase agreement with the plant in 2024, and the latest reporting suggests that construction and regulatory procedures for an actual restart are moving forward. The case is drawing attention as an example of AI infrastructure demand shaping energy policy itself.

Chainwash Attack Neutralizes 94% of Diffusion LM Watermarks

A paper published on arXiv, "Chainwash," reports a technique that breaks the statistical watermarks embedded in diffusion language models (Diffusion LMs) using multi-stage rewriting attacks.

The target is the watermarking method of Gloaguen et al. applied to LLaDA 8B Instruct, which normally boasts a detection rate of over 99%. However, the research team used four open-weight models (1.5B to 8B) and applied five rewriting styles (paraphrasing, humanizing, simplification, academic tone, and summary expansion) up to five times in a row; the watermark detection rate then plummeted from 87.9% to 4.86%, and 94.76% of the originally watermarked texts became undetectable.

After just three rounds of rewriting, detection scores dropped by 86%, shifting from the watermarked threshold toward the unwatermarked distribution. Repeated rewriting proved to be far more powerful than a single rewrite. For AI-generated content detection and provenance technology, these results pose a serious challenge.

Colopl Releases Creator Protection Tool "CCP" for Free

Colopl, a company active in everything from game development to AI applications, has released "COLOPL Contents Protector (CCP)," a free app that protects creators' works from unauthorized training by generative AI.

In an ITmedia interview, CIO Kenta Sugai (Senior Executive Officer) and Tsuyoshi Kudo, who led development, explained their seemingly contradictory position — developing a protection tool while being an AI-promoting company — starting from the creator's perspective that "no illustrator wants to put watermarks on their work." Their stance of actively embracing AI while also exploring ways to protect works from it is drawing attention as a responsible approach for an AI company.

DDR6 Delayed Again, Commercial Products Slip to 2028 — A Bottleneck for AI Hardware

Commercialization of the next-generation memory standard DDR6 has been delayed yet again, and it is now expected to slip to 2028. As discussed in Reddit's LocalLLaMA community, DDR6 was predicted to arrive in 2026 just a few years ago, but the schedule has fallen far behind.

The same thread has also seen lively debate over the idea that "hardware is the only moat (the source of competitive advantage)." With LLMs alone beginning to show their theoretical limits and open source likely to catch up in the end, some argue that securing the hardware and compute resources that underpin inference capabilities will become the real differentiator. As the tendency to prioritize data center GPUs strengthens, soaring consumer hardware prices are another cause for concern.

Paper Flags the "Authorization Propagation" Problem in Multi-Agent AI

"Authorization Propagation in Multi-Agent AI Systems," published on arXiv, presents a different kind of problem for agentic AI security discussions, which tend to focus heavily on prompt injection: the propagation of authorization.

When multiple AI agents work together to retrieve, delegate, and integrate data, maintaining permissions across non-human principals (agents) is more than conventional access controls such as RBAC and ABAC can handle. The paper formalizes this as "authorization propagation" and identifies three sub-problems: transitive delegation, aggregation inference, and temporal validity. Preliminary empirical findings from production enterprise AI platforms are also reported, showing that the problem arises during normal operation, not just under adversarial attack.

Sley: The First Native AI Programming Language Arrives as Open Source

"Sley" has been released on GitHub. It bills itself as a "native AI programming language" designed on the assumption that AI agents write and execute code directly, and it has also been attracting attention on Hacker News.

Conventional programming languages are designed assuming that humans write them, whereas Sley is said to adopt syntax and semantics optimized for LLM generation patterns. As ecosystems where AI agents autonomously generate and execute code continue to expand, attention is focused on what advantages AI-native language design can deliver, from the dual perspectives of practicality and safety.