US Government Approves Limited Re-Release of Anthropic's "Mythos"
On June 26, the US government approved the limited re-release of Anthropic's powerful AI model "Mythos" to "trusted US organizations." The news drew significant attention, earning 310 points and 319 comments on Hacker News.
According to reports from Reuters and Semafor, the model will be made available only to select US companies. Global general availability remains on hold, effectively restricting access to the most powerful AI models to a limited set of organizations within the United States.
Criticism: Is This an AI Model "Monopoly"?
Following the limited release of Mythos, debate has intensified on Hacker News over whether the US government is trying to monopolize AI as a superpower. Commenters pointed out that the release of OpenAI's GPT-5.6 (Sol) was also delayed by government order, and that the global rollout of Anthropic's Claude (Fable 5) likewise remains on hold.
Some argue that restricting the most powerful AI models to "trusted US organizations" has effectively created an AI monopoly. The situation raises, once again, the question of how to strike a balance between national security concerns and the democratization of technology.
Behind Treasury Secretary Bessent's AI Concerns: "Threats to Payment Rails"
According to another Semafor report, Treasury Secretary Scott Bessent became deeply involved in AI issues after warnings about AI-derived threats to the Federal Reserve's payment rails. The possibility that advances in AI could undermine the safety of financial infrastructure appears to be accelerating the government's move toward AI regulation.
Exam Cheating with "AI Glasses" Spreads Across Asia
Social issues around AI wearables are also emerging. According to RNZ, a growing number of students—particularly in Asia—are using AI glasses to cheat on exams. Asia's exam-focused education culture has made the region "ground zero" for the phenomenon, and technical countermeasures in educational settings are now an urgent priority.
"Agentjacking": Hijacking AI Coding Agents
A new attack technique targeting AI coding agents, dubbed "Agentjacking," is drawing attention. According to a Qiita article based on reporting by InfoSecurity Magazine, the attack—which tricks agents into executing arbitrary commands through fake error reports—achieved an 85% success rate in experiments. As development workflows that delegate everything from writing code to investigating and fixing errors to AI agents become commonplace, this new security threat goes beyond prompt injection and warrants vigilance.
Major CUDA Performance Improvements Merged into llama.cpp (b9820)
The open-source LLM inference engine llama.cpp has merged major performance improvements to its CUDA backend. By reducing synchronization between tokens and introducing asynchronous CPU-to-CUDA copies, the change is expected to improve inference throughput. The benefits could extend to other backends such as Vulkan, and the news has become a hot topic in the local LLM community.
---