Anthropic Shuts Down Newest AI Models Under White House Pressure

The White House's weekend order forcing Anthropic to cut off international access to its newest AI models landed on a day the U.S. was celebrating two sports championships — but inside the AI industry, it marked the clearest signal yet that frontier models sit squarely within reach of executive power.

Security

A China-linked espionage group spent more than a year inside North American medical, academic, and military research networks, exfiltrating sensitive communications by abusing Google Workspace mail-routing rules. The initial foothold came through a backdoor on REDCap research servers; from there, attackers quietly configured Workspace forwarding rules to siphon defense and research email without triggering obvious alerts. Fourteen-plus months of silent dwell time combined with use of a trusted productivity layer for exfiltration are hallmarks of a well-resourced state-sponsored collection operation.

North Korean operators are simultaneously running two active campaigns through the Contagious Interview cluster (also known as Famous Chollima and Void Dokkaebi), converting developer tools into malware delivery channels via fake job interview lures. When the people building software become the primary attack surface, the downstream risk compounds: a compromised developer workstation is a foothold inside every system that developer maintains.

LiteLLM, a widely deployed open-source AI API gateway, has a three-vulnerability chain that lets a default low-privilege account escalate to full admin and achieve remote code execution, per Obsidian Security. The broader implication: organizations increasingly route all AI traffic through proxy middleware like this, which makes a single compromise a pivot into every model backend the gateway serves. AI gateway security is now a real attack surface, not a hypothetical one.

AMD has quietly removed Transparent Secure Memory Encryption (TSME) from its consumer CPU line, Ars Technica reports. Users are angry not just at the removal — TSME costs essentially no performance while protecting against cold-boot and physical-access attacks — but at the apparent covertness of the decision. Silently stripping a security feature from consumer silicon is a meaningful trust regression.

Two closing supply-chain threads: Cisco has patched CVE-2026-20262 in Catalyst SD-WAN Manager, a root-privilege-escalation zero-day under active exploitation, closing the remediation arc on a flaw previously reported here. The Awesome Motive CDN compromise has widened to include OptinMonster, TrustPulse, and PushEngage, extending the WordPress supply-chain blast radius covered in recent days. And on the enforcement front, the DOJ announced the first seizures under the TAKE IT DOWN Act, taking down CFAKE.com and SOCFAKE.com, which hosted nonconsensual AI-generated nude imagery — a new legal mechanism making its enforcement debut.

AI

The dominant story is the Trump administration's order to Anthropic to immediately block all foreign access to Fable 5 and Mythos 5, its newest and most capable models. The order arrived at 5:21 PM on a Friday, forcing a weekend shutdown. The Verge's reporting and TechCrunch's analysis both conclude this was not primarily about a jailbreak or security exploit — the framing initially floated. Anthropic says it had little choice and is reportedly in talks with the administration toward some accord.

The sovereignty read is the sharper edge. An American AI company, operating under U.S. law, can have its global product reach severed at a weekend's notice by executive action. The Verge notes this immediately makes the case for non-American AI alternatives — precisely the competitive outcome that export-control hawks typically want to prevent. Whatever the accord looks like, the template for government interference in frontier AI has now been demonstrated.

Meta, meanwhile, is pursuing AI reach through a different route: AI Mode on Facebook is a generative search layer that draws on public posts across Meta's platforms in real time. It's a catch-up play on AI-native search, but the data provenance angle is material — public Facebook posts are now live retrieval signal feeding AI-generated results.

Tech

Fox is acquiring Roku for $22 billion, taking control of the streaming OS, hardware, and FAST channel stack that sits in more than 100 million homes globally. This is a bet on the living room as an advertising platform, with Fox gaining direct control over the surface that mediates how cord-cutters find content. A media company owning the OS layer — and with it the recommendation engine, ad inventory, and viewer data pipeline — has significant implications for how streaming shakes out.

SpaceX has gone public, ending years of speculation. The company's financials are now subject to public disclosure for the first time, which will be illuminating for an industry that has largely taken SpaceX's dominance on faith rather than visible unit economics.

Nvidia is raising at least $25 billion in bonds — its first debt issuance since 2021 — testing investor appetite for AI sector exposure at scale. At the same time, Microsoft is turning to AWS to handle AI compute overflow from GitHub's capacity crunch, a telling sign that even hyperscalers are straining under the AI build-out.

Xbox's "reset" is becoming a rout: Ninja Theory (Hellblade) and Compulsion Games (South of Midnight) are both being shut down alongside executive departures. Microsoft's first-party studio footprint is contracting sharply.

The UK is moving to ban social media for users under 16 with possible overnight access curfews, joining a growing list of countries adopting age-restriction frameworks. Enforcement remains the unresolved variable; critics note VPNs trivially circumvent these measures.

The Anthropic episode crystallizes the defining tension of this AI moment: the same government pushing AI dominance as a national-security priority can also reach in and shut the product off on a weekend.

Also yesterday

• Palo Alto Confirms Active Exploitation of GlobalProtect VPN Auth Bypass — CVE-2026-0257
• WordPress Plugin Supply-Chain Attack Plants Hidden Admin Backdoors
• Microsoft 365 Copilot 'SearchLeak' Flaw Enables 1-Click Enterprise Data Theft