Anthropic's Export-Control Blackout Exposes the Kill Switch at the Heart of American AI

The day's biggest story isn't a vulnerability or a model launch — it's an export control order that briefly made Anthropic's AI inaccessible to its own employees, and the question it forced into the open: who actually controls access to American AI?

Security

A massive credential dump exposed login credentials for thousands of enterprise networks, with named victims including Oracle, Lenovo, FedEx, a NATO contractor, and — in a particularly awkward detail — Fortinet itself. TechCrunch's reporting attributes the campaign to an alleged Russian-speaking group exploiting previously known passwords across Fortinet firewalls and VPNs. The method is embarrassingly low-tech: reused or weak credentials on perimeter devices. If you run Fortinet gear, this is the moment to audit credential hygiene across every management interface. The fact that Fortinet's own infrastructure appears in the dump is not a good sign for vendor trust.

Separately, Microsoft confirmed a Defender zero-day codenamed RoguePlanet, now assigned CVE-2026-50656 with a CVSS of 7.8. A patch is in development but hasn't shipped. Details on active exploitation haven't been released, but a 7.8 in Defender — the thing meant to protect you — warrants close attention. No workaround has been published; watch the MSRC advisory.

On the privacy side, Google announced it will begin using IP addresses from UK, EEA, and Swiss users for ad measurement and personalization starting August 3, 2026. The timing is notable: Google itself argued for years that IP addresses were insufficient to identify individuals — a position it apparently no longer needs. This lands as the UK's ICO is revising consent rules. Expect legal challenges well before the August date.

AI

The most consequential AI story right now isn't a benchmark or a product launch. The Trump administration's export control order forced Anthropic to abruptly cut off access to its newest models for all foreign nationals — including users physically inside the US and Anthropic's own employees. The company spent days fighting to restore service. The episode made viscerally clear something that had been theoretical: American export controls can act as an instant kill switch on AI access, regardless of where users are located or whether they work for the company that built the model.

That anxiety is now explicit at the diplomatic level. At the G7, French President Macron and Indian PM Modi both raised the concern that the US could cut off access to American AI overnight. They want the capabilities; they don't want the dependency. The Anthropic blackout wasn't hypothetical anymore — it had just happened. Expect this to accelerate European and Indian investment in domestic AI infrastructure, not because they'll catch up quickly, but because the political cost of dependence just became undeniable.

Against that backdrop, Noam Shazeer — transformer co-author and one of the architects of modern LLMs — has joined OpenAI. Shazeer had been at Character.AI before Google's acquisition. His move to OpenAI is a significant talent signal, particularly in a moment when fundamental architectural expertise matters more than headcount.

On safety: ChatGPT spontaneously generated sexual violence and snuff imagery from a viral prompt, per a Mindgard report. The structural issue is familiar — content policies that fail on edge-case prompt combinations — but the failure mode here is severe. OpenAI has not publicly responded.

Pew Research data is worth sitting with: only 16% of Americans think AI will have a positive impact on society, even as 49% report using chatbots at least occasionally, and 63% think the technology is advancing too quickly. Usage and trust have fully decoupled. People are using the tools while remaining deeply skeptical of the trajectory — a dynamic that should be legible to anyone following the Anthropic export control story.

Tech

Apple is raising prices. Tim Cook told the Wall Street Journal that RAM and storage cost increases are "unsustainable" and that price increases are unavoidable. This is partly a memory shortage story and partly a tariff story — AI-driven memory demand is compressing supply across the board. The next iPhone cycle will cost more; enterprise procurement math gets complicated quickly.

Midjourney made an unexpected hardware pivot. CEO David Holz unveiled the Midjourney Scanner, a full-body ultrasonic CT device under the Midjourney Medical brand. Holz acknowledged it's a departure from the "cat pictures" product. Whether Midjourney can execute on FDA-regulated medical hardware is a separate question from whether the AI imaging underpinnings are compelling. The clinical validation timeline is not yet clear.

Snap unveiled its long-awaited AR glasses — and the stock tanked. The hardware is real, but pricing is prohibitive and Snap's path to mass-market AR remains unconvincing to investors. The company hasn't made the core aspiration legible in a way that justifies the capital requirements.

Tesco is migrating 40,000 server workloads off VMware, citing a roughly 175% price increase after Broadcom's acquisition and what Tesco's UK court filings call "abusive conduct." Tesco is large enough to litigate and large enough to migrate. Most organizations aren't. The Broadcom/VMware pricing story continues to be one of the more consequential enterprise infrastructure stories playing out mostly in court filings and quiet platform exits.

Credential hygiene on perimeter devices remains the failure mode nobody fixes until it's too late — and today Fortinet itself is on the victim list.

Also yesterday

• Malicious JetBrains Marketplace Plugins Are Stealing AI API Keys
• Google Vertex AI SDK Flaw Let Attackers Run Code in Google's Infrastructure via Bucket Squatting
• CISA Flags Actively Exploited Joomla JCE Plugin for PHP Code Execution