Nadella Warns Enterprises Off Proprietary AI as Russia Targets Home Routers

Two warnings defined the day: one from Washington about the router in your closet, one from Redmond about the AI vendors in your contract — and the more surprising of the two came from Microsoft.

Security

CISA, NSA, and intelligence partners from the UK, Canada, Australia, and New Zealand jointly published an advisory warning that Russian state-sponsored actors are actively compromising consumer and SOHO routers to build residential proxy infrastructure. The technique lets attackers blend malicious traffic through IP addresses that appear to be ordinary homes, making geo-fencing and IP-reputation lists ineffective — the traffic looks like it originates in Minneapolis or Manchester, not Moscow. The advisory recommends the standard hardening checklist: update firmware, disable remote management, rotate default credentials. The underlying problem it doesn't solve is that most consumer routers have short firmware support windows, and many units currently deployed are past end-of-life with no patch path available. Residential proxy abuse has been a persistent feature of Russian state operations for years; this advisory signals the targeting is broadening in scope.

That backdrop arrived the same day Nihon Kotsu, Japan's largest taxi operator, announced it was forced to shut down portions of its IT infrastructure following a cyberattack. The company hasn't disclosed the attack type, entry vector, or extent of data exposure. Operational shutdowns at a company running Japan's largest taxi fleet translate directly into service disruption across a major transportation network. Details are still emerging, but the pattern — infrastructure operator goes dark, then issues a brief, vague disclosure — is a familiar ransomware signature.

AI

The sharpest signal of the day came from Satya Nadella. The Microsoft CEO published a post explicitly warning enterprises against locking into proprietary AI models, naming Anthropic and OpenAI specifically. This is a striking move: Microsoft is one of OpenAI's largest backers, having invested heavily and built Microsoft 365 Copilot on OpenAI models. Nadella framing that relationship as a vendor lock-in risk for enterprises either reflects a genuine shift toward open models — which Microsoft has been quietly expanding on Azure — or a positioning play designed to route enterprise AI spend through Azure's model marketplace, where Microsoft profits regardless of which model wins. Either reading has implications: enterprise procurement teams will be citing this statement in vendor negotiations before the week is out.

MIT researchers unveiled a method for detecting whether an AI model was trained on child sexual abuse material without requiring the model to surface any of it during evaluation. The prior detection problem was circular: proving contamination required eliciting contaminated outputs, which creates evidence that is itself illegal to produce or possess. The MIT approach uses activation probing to find the forensic signal in model weights without surfacing the content. This gives AI safety auditors, hosting platforms, and regulators a legally viable screening path that didn't exist before.

Nous Research, makers of the Hermes open-weight model family, is in talks to raise at least $75M at a $1.5B valuation, led by Robot with participation from USV. Open-weight labs attracting pre-IPO-scale investment on the same day a major proprietary AI backer urges diversification away from proprietary models isn't coincidental — it reflects where a segment of enterprise buyers is already hedging.

Tech

Apple released public betas for iOS 27, macOS 27 Golden Gate, and watchOS 27, opening the fall OS lineup to public testing. The centerpiece is Siri AI, Apple's long-delayed on-device AI overhaul, and early testers report it delivers. The watchOS 27 build is drawing particular attention for making the Watch feel genuinely useful in ways that have eluded it since launch. macOS 27 ships with a more subdued Liquid Glass aesthetic — a clear response to developer preview feedback that the original effect was visually overwhelming. Public betas mean anyone can test this without a developer account.

Samsung issued a data ultimatum: users who decline to consent to their Samsung Health data being used for AI model training will have that data deleted. There is no middle path — accept the training consent or lose the data. In a health data context this is an aggressive posture, and it is almost certainly headed for regulatory scrutiny in the EU, where health information sits in a GDPR special category that requires explicit, granular consent and cannot legally be bundled with service-continuation ultimatums.

Apple disclosed that a former employee allegedly exploited a post-departure access bug to exfiltrate confidential files from Apple's network after leaving to join OpenAI. Apple called the vulnerability "rare," which is notable primarily for what it implies about offboarding gaps when employees move directly to frontier AI competitors. The case has obvious trade-secret dimensions and adds another data point to the increasingly fraught talent pipeline between established tech companies and AI labs.

OnePlus and Oppo are reportedly preparing to exit the US and European markets in the coming days. Trade pressure on Chinese hardware, inability to scale at the premium tier without deep local investment, and a narrowing value-segment window together make both markets untenable. Enthusiast Android buyers lose one of the last major alternatives to Samsung and Apple at the near-premium price point.

SpaceX is targeting Starship's 13th integrated test flight in the days ahead, pushing the vehicle to higher structural pressure and attempting to deploy functional Starlink satellites from orbit for the first time — a real payload delivery rather than a pure hardware stress test.

Infrastructure under pressure from state actors and ransomware groups, the open-versus-closed AI debate gaining a prominent new voice, and Apple's fall platform story now in public hands.

Also yesterday

• CISA Flags Two Maximum-Severity Joomla Zero-Days Under Active Exploitation
• Three Live Microsoft 365 Evilginx Phishing Ops Exposed by Attacker's Own Blunder
• Grok CLI Silently Uploaded Entire Home Directories to xAI Cloud Storage
• Malicious Jscrambler npm Package Installs Infostealer on ~1,500 Systems