The EU Reworks Its AI Act — Correcting Regulatory Overreach?

The EU has rewritten the AI Act, the pioneering AI regulation it adopted ahead of the rest of the world, and the move is making waves. According to [The Parliament Magazine](https://www.theparliamentmagazine.eu/news/article/why-the-eu-rewroteits-landmark-ai-law), the revision is seen as a response to concerns that the strictness of the regulation has been undermining industrial competitiveness.

The AI Act took a "risk-based" regulatory approach, but critics have long pointed out that, in practice, the burden on companies proved heavier than expected. The key question going forward is how much of that practical burden this rewrite will actually ease.

Microsoft Warns: A Vulnerability That Lets AI Agents Be Hijacked With a "Single Page"

Microsoft's security blog has disclosed an attack technique dubbed [AutoJack](https://www.microsoft.com/en-us/security/blog/2026/06/18/autojack-single-page-rce-host-running-ai-agent/). The details are alarming: merely by having an AI agent browse a web page, an attacker can achieve remote code execution (RCE) on the host machine.

As use cases multiply in which AI agents access external content through a browser, this kind of attack is becoming a realistic threat. If your design lets an agent crawl arbitrary URLs, it may be time to re-examine your sandboxing and content filtering.

Russia's AI Sovereignty Ambitions and the "Chip Problem"

[Time](https://time.com/article/2026/06/18/russia-ai-putin-chip-us-china/) reports that while Russia aims to establish AI sovereignty, semiconductor procurement has become its biggest obstacle.

The Putin regime has placed AI at the core of its national strategy, but U.S. and Chinese export controls have made it difficult to secure high-performance AI chips. The prevailing view is that unless the country can build a semiconductor supply chain of its own, its drive for self-reliant AI development will run into a hard limit.

Local LLM Frontlines: GLM 5.2 Inference Speeds and What Gemma 4 Can Do

Local LLM talk is buzzing in the LocalLLaMA community.

First up, GLM 5.2: users have been sharing their inference speeds in a [Reddit thread](https://www.reddit.com/r/LocalLLaMA/comments/1ub69xu/glm_52_what_speeds_are_we_getting_locally/). According to the reports, a 6x RTX 3090 setup running IQ2_M quantization with a 90K context achieved generation speeds of about 7.8 tokens/sec. Prompt processing came in at around 40 tokens/sec.

As for [Gemma 4 26B](https://www.reddit.com/r/LocalLLaMA/comments/1ub4ods/gemma_4_26b_a4b_is_genuinely_the_best_model_i/), some rate it below Qwen 3.5/3.6 on coding tasks, but others call it among the very best for language learning and scientific queries (medicine, biology, and biochemistry). As small MoE models in the same 20–30B class, the two are drawing close comparisons.

Meanwhile, a [Japanese article on Zenn](https://zenn.dev/aiwatch_jp/articles/gemma4-e4b-qat-local-hands-on) put Gemma 4 E4B (QAT version) to the test on real hardware. Despite being a 7GB model, it reportedly solved competition math, read through 27K tokens of code, and even flagged a calculation error on an invoice.

Why Enterprise AI Is "Stuck"

A [FastCompany analysis](https://www.fastcompany.com/91555415/real-reason-enterprise-ai-stuck) digs into why AI adoption in enterprises has fallen short of expectations. The argument: rather than technical problems, the fundamental challenges lie in organizational data foundations, process redesign, and how ROI is measured.

The WSJ, for its part, reports on [the gap between the market's AI exuberance and political reality](https://www.wsj.com/tech/ai/the-markets-ai-fanfare-is-running-into-a-harsh-political-reality-b919829b), noting that uncertainty in the regulatory environment is casting a shadow over investment decisions.

---