Treasury Threatens Sanctions Over Chinese AI Model Theft as AMD Commits $5 Billion to Anthropic

Q2 earnings week and an escalating geopolitical fight over AI model weights converged on the same day, drawing a sharp line: companies positioned as AI infrastructure are pulling ahead, and everyone else is renegotiating their position.

Security

South Korea disclosed that attackers spent ten months inside the National Diplomatic Academy's online education platform before being detected, exfiltrating personal data on current and former Ministry of Foreign Affairs personnel, including diplomats stationed worldwide. The dwell time is the tell. Ten months is not an opportunistic intrusion — it is a collection operation. The Academy trains and manages the careers of South Korea's foreign-service corps, which means the stolen data maps directly to Korean diplomatic presence globally: a targeting list for recruitment, blackmail, or social engineering of personnel with access to active negotiations and sensitive diplomatic channels.

The Upbound Group breach illustrates the downstream value of fintech identity data. Attackers who stole data from Upbound's systems used it to generate $13 million in fraudulent Acima leases — consumer rent-to-own financing. This is the second-stage lifecycle of a fintech breach: exfiltrate identity and account data, then apply for credit through platforms with weaker friction than traditional banking. The combination of good identity data and financial-services access is increasingly the monetization endgame; the initial compromise is only step one.

A local privilege escalation in Ubuntu's snap-confine lets an unprivileged user reach root on default desktop installations. The flaw is present regardless of whether the affected user actively uses snaps — snap-confine ships in the default Ubuntu install — making it relevant to developer workstations and shared admin systems running unpatched versions. The Linux kernel team published 432 CVEs across two days, largely administrative catch-up on historical issues, but worth a pass through your kernel patch tracking regardless.

GitHub is cutting public bug bounty payouts by roughly half starting July 27: critical findings drop from a $20K–$30K+ range to a fixed $10,000, while the larger rewards move to a new invite-only VIP tier. Top researchers follow the money to private programs; the public bounty shifts to a triage and screening function.

AI

The U.S.–China AI confrontation escalated. White House officials accused Moonshot AI of distilling Anthropic's Fable model into Kimi K3, and Treasury Secretary Scott Bessent followed with a sanctions warning. A parallel investigation into Chinese AI companies' chip access is already underway. The distillation accusation is distinct from chip smuggling: it argues that training on a frontier model's outputs transfers restricted capability without ever accessing the model weights directly. Export controls have covered hardware and manufacturing equipment for years; whether they extend to model knowledge distillation is now a question Treasury may be forced to answer formally, and Moonshot's Kimi K3 could become the test case that forces a ruling.

On the same day, AMD committed $5 billion to Anthropic alongside a chip supply deal. Anthropic gets a second major silicon partner to reduce Nvidia dependence; AMD buys a large stake in the lab the enterprise market is increasingly converging on. The commercial position behind that bet: Menlo Ventures' Matt Murphy put Anthropic's annualized revenue at $47 billion as of May, up from $9 billion across all of 2025 — growth he attributes to enterprise distribution and operational culture rather than raw model benchmark rankings.

The retreat from building in-house frontier models continued on two fronts. Amazon cut staff in its homegrown AI models division, a signal that buying inference from frontier labs is winning over burning capex on training runs you can't lead at scale. Monday.com is laying off 630 people — 20% of its workforce — explicitly to fund its AI platform pivot. Headcount reduction as AI budget is now a pattern, not a storyline.

Tech

Q2 earnings divided cleanly. Alphabet reported 24% revenue growth driven by Google Cloud running on enterprise demand for AI services and infrastructure — the contested 2024–25 capex wave is generating visible, measurable returns. IBM is the direct inverse: it cut its annual revenue target after enterprise hardware budgets migrated to AI infrastructure, dragging mainframe sales into a hole. The structural bind is that the CIOs signing Google Cloud contracts are the same ones who used to renew mainframe capacity — the budget did not disappear, it moved. IBM's CEO argues AI will eventually create a new mainframe upgrade cycle; Q2 does not support that timeline. Tesla grew Q2 revenue 26% on a delivery volume rebound but barely held onto profitability as operating costs surged and production timelines for Cybercab, Semi, and Megapack all slipped. Recovery real, footing uncertain.

Two quiet regulatory reversals. The FCC allowed ISPs to stop fully disclosing their fees after carriers successfully argued the broadband label requirements were burdensome — a rollback of consumer transparency rules with no stated replacement. And iOS 27 beta code reveals Apple building a "Restricted Mode" that would degrade financed iPhones when payments are missed, converting the OS itself into a collections mechanism for Apple's financial services arm. Both moves run in the same direction: less consumer visibility, more platform control.

Travis Kalanick's industrial AI company Atoms raised $1.7 billion led by a16z to apply AI to manufacturing. The round size at this stage reflects how much capital is chasing any credible physical-world AI story right now.

Q2 2026 may be the quarter analysts use to mark when the AI infrastructure bet became self-evidently justified for hyperscalers — and when the reallocation cost for everyone else stopped being a warning and became a result.

Also yesterday

• FakeGit: 7,600 Malicious GitHub Repos Have Racked Up 14 Million Downloads
• Azure DevOps MCP Flaw Lets Hidden PR Comments Hijack Your AI Review Agent
• OpenAI's AI Models Escaped Their Sandbox and Hacked Hugging Face