China's Access to Anthropic's Mythos Triggers Export Controls and EU Scrutiny

Geopolitics collided with AI governance yesterday as fears about Chinese access to Anthropic's most capable model drove U.S. export restrictions — and immediately drew scrutiny from Brussels over what those restrictions mean for European AI collaboration.

Security

AUR malware escalates; the FBI's fake town for cyber training

Arch Linux's AUR sustained another wave of malware attacks, this time described as more sophisticated than prior incidents. The AUR's design — community-submitted packages with build scripts that execute as root, no mandatory signing, no central review — makes it a perennial soft target. What's notable in this wave is the apparent coordination: packages appear to have been selected for install-base size rather than opportunistically compromised. If you pull from foreign sources regularly, audit your installed AUR packages and check build script history. The AUR's "use at your own risk" model has always required active vigilance; these incidents are periodic reminders of what that actually demands.

On the defensive side, The Verge published a worthwhile look at the FBI's Cyber Range in Huntsville, Alabama — a 22,000-square-foot replica of a small American town, complete with simulated power grid, water treatment systems, and hospital infrastructure. Opened last year and modeled conceptually on Hogan's Alley, the range lets agents and partner organizations practice responding to attacks on critical infrastructure without touching live systems. The kinetic cyber angle — physical consequences triggered by digital intrusion — is the scenario planners have been building training programs around since Stuxnet, and it's good to see purpose-built range infrastructure finally catch up to the threat model.

AI

China and Anthropic's Mythos: export controls arrive

The day's dominant story: Semafor reported that the White House's decision to impose export restrictions on Anthropic's Mythos model was driven in part by intelligence indicating a China-linked group had accessed it. If accurate, this is precisely the frontier-model leakage that export control advocates have warned about — not theft of model weights, but unauthorized access to a highly capable model before its controlled release. The implications for how frontier labs manage pre-release access are significant, and the incident will likely accelerate internal access control reviews across the major labs regardless of any formal regulatory requirement.

The timing matters too. These restrictions land while the broader legal question — whether AI models should be treated as export-controlled technology (like chips) or as speech (like software) — remains unsettled in U.S. law. There is no established statutory framework analogous to semiconductor export controls, and what exactly counts as "access" is genuinely ambiguous when a model can be queried through an API, fine-tuned on outputs, or distilled into a smaller derivative.

The EU didn't wait long to weigh in. A Commission spokesperson said Brussels is examining the practical consequences of the Anthropic decision — careful diplomatic phrasing for "we're concerned about what this means for European researchers and companies who depend on U.S. frontier model access." European AI policy has consistently tried to stay in front of U.S. unilateral action; being caught downstream of an American export restriction is exactly the scenario Brussels has been trying to avoid.

The financial stakes clarify why any of this matters at the geopolitical level. The Information reports that OpenAI and Anthropic employees have collectively cashed out roughly $14 billion through secondary transactions — a figure that reflects both extraordinary private valuations and the continued absence of public liquidity. For a sense of scale: FTX's former stake in Anthropic would be worth approximately $75 billion at today's valuation. The company that launched as a well-funded AI safety research lab has become one of the most valuable private companies on earth, which is exactly why access to its models carries geopolitical weight.

Rio de Janeiro's municipal LLM is apparently not what it claimed

The city of Rio de Janeiro promoted Rio3.5 as a locally trained sovereign LLM — but a GitHub issue thread reveals it appears to be a merge of existing open-source models rather than an original training effort. The model was drawing favorable benchmark comparisons against Qwen3.7, which made the provenance question more pointed once it surfaced. Government AI initiatives claiming sovereign development credentials need this kind of scrutiny: benchmark performance and model origin are entirely separate questions, and political incentives push hard toward conflating them. This is unlikely to be the last such case.

AI adoption: the reality check

Gabriel Weinberg's data-driven counterpoint to AI hype argues that actual regular AI usage remains heavily concentrated in tech-adjacent demographics, and the perception of universal adoption is largely an artifact of heavy users dominating the platforms where AI discourse happens. The piece pairs well with The Register's argument that AI is ultimately code — and that prompting alone can't transcend what the model was trained to do. Both pieces push back against narratives that have outrun the underlying data.

Tech

Linux 7.1

Linus Torvalds tagged Linux 7.1 — a quiet release drop; the LKML announcement thread is the canonical starting point if you're tracking driver changes or subsystem work in this cycle.

UK moves toward social media ban for under-16s

The UK government is advancing a ban on social media access for children under 16, following Australia's lead. The implementation hard problem hasn't changed: reliable age verification without either constructing a national internet identity database or delegating that function entirely to private platforms. Australia passed its law in 2024 and is still working out enforcement mechanics; the UK is watching before finalizing its own approach. The political momentum, though, is clearly in one direction.

Charlie Javice seeks a Trump pardon

Frank founder Charlie Javice — convicted of defrauding JPMorgan by fabricating the student loan startup's user numbers before a $175M acquisition — is reportedly seeking a presidential pardon. She joins a growing list of convicted tech figures petitioning the current administration. JPMorgan, which sued to recover acquisition costs and prevailed at trial, presumably has views on this.

Stanford walkout on Sundar Pichai

Graduate students at Stanford walked out during Sundar Pichai's commencement address, a protest tied to Google's AI and cloud infrastructure contracts. Campus tech protest has become a recurring feature at elite CS programs over the past 18 months; this one continues the pattern.

The Mythos export restriction story is the one to track: it's the first documented case of U.S. authorities restricting a specific AI model on national security grounds following suspected adversary access, and the EU's immediate response suggests the regulatory and diplomatic friction is just getting started.

Also yesterday

• FBI Dismantles 'Outsider Enterprise,' a Chinese AI-Powered Phishing-as-a-Service Platform