Anthropic Significantly Expands Project Glasswing — Now Covering Critical Infrastructure in 15+ Countries
Anthropic has significantly expanded its security program, Project Glasswing. The initiative has grown from roughly 50 participating organizations to around 150, with coverage now extending to more than 15 countries. Using Claude Mythos Preview to scan codebases for vulnerabilities, the program has uncovered more than 10,000 high-severity and critical security defects to date.
The expansion includes many organizations in critical infrastructure sectors such as power, water, healthcare, telecommunications, and hardware. Many of the partners are organizations whose compromise could affect more than 100 million people. Vendors and open source maintainers are also newly covered, aiming to raise security standards across a broad swath of the supply chain.
As AI reshapes the assumptions behind cybersecurity, this initiative stands out as one of the most concrete implementations of "defense with AI." That said, whether the rollout to 150 organizations actually translates into fewer vulnerabilities will depend on the quality and continuity of execution.
OpenAI Calls for Global Action on AI Safety for Young People
OpenAI has issued a statement calling for global efforts to protect young people in the age of AI. The company advocates building youth protection frameworks in coordination with governments around the world, including the establishment of a dedicated AI Safety Institute.
AI's impact on young people spans many areas, from content appropriateness to mental health effects. OpenAI is showing a willingness to take part in shaping the regulatory framework itself, but at the same time it faces calls for concrete actions such as age limits and stronger safety features for its own products. Whether the statement is backed by effective measures will be the key point to watch going forward.
Hugging Face Releases Local Computer Use Agent "Holo3.1"
A post on the Hugging Face blog introduced Holo3.1, a locally running Computer Use Agent developed by Hcompany. The agent runs entirely in a local environment without relying on the cloud and can autonomously perform on-screen operations.
Computer Use Agents have mostly been offered as cloud services so far — Anthropic's Claude and OpenAI's Operator being notable examples — but Holo3.1's distinguishing feature is that it runs fully locally, preserving privacy. It holds promise for edge devices and enterprise environments, though whether its accuracy has reached a practical level will need separate verification.
"AI Doesn't Have ROI" — The Debate Over Returns Heats Up
An article titled "AI Doesn't Have ROI" from Where's Your Ed At is drawing attention on Hacker News. Its argument: despite massive investments in AI, the actual return on investment (ROI) remains hard to see.
In the AI boom, companies are racing to adopt AI features, but at many of them "using AI" has become an end in itself, with no numbers presented for actual productivity gains or cost savings — a pointed critique. While generative AI use cases keep emerging one after another, few companies can yet answer the basic question of "how much money did this actually make?" This debate looks set to become an important bellwether for the direction of future AI investment.
Interconnects Author Leaves Ai2 — What It Means for Open Source AI
The author of the AI newsletter Interconnects, who worked on the Olmo model at the Allen Institute for AI (Ai2), has announced their departure. Looking back on their experience developing open source AI models at Ai2, they touched on the possibility of making an impact along a path other than the frontier-model performance race.
For the open source AI community, losing a core developer is a cause for concern. That said, contributions to the community are expected to continue after the departure — if anything, new forms of contribution may be on the way.
Polish Prime Minister Advocates "Technological Sovereignty" to Counter AI Risks
Poland's prime minister has argued that "technological sovereignty" should be made a top political priority in order to counter AI risks. According to Bloomberg, the prime minister indicated that technological self-reliance is essential to protect the country's security and economy.
With interest in AI sovereignty growing across Europe, Poland's declaration offers a useful reference point for other Central and Eastern European countries. What "technological sovereignty" would actually entail — nurturing domestic AI or tightening regulation — remains unclear, however.
Bonsai Image 4B — An Ultra-Compact Image Generation Model That Runs in 0.93GB
Bonsai Image 4B is attracting attention in Reddit's LocalLLaMA community. It's a Diffusion Transformer capable of generating images with a footprint of just 0.93GB for the 1-bit quantized version, and 1.21GB even for the Ternary version.
Shrinking on-device image generation down to this size is a major step forward for practical edge AI. Real-time image generation on smartphones and other small devices is starting to look realistic.
llama.cpp Gains a Thinking Mode Toggle
llama.cpp Pull Request #23434 adds a Thinking mode toggle switch to the UI. The feature lets you adjust how much reasoning goes into inference — turning thinking off for simple questions and on for complex reasoning.
It's an update that further improves the open source LLM experience, and it's especially useful for balancing resource efficiency and accuracy in local setups.