TechWithAdit (TWA)

Microsoft cut 4,800 people today — Xbox is losing 20% of its workforce and four gaming studios are being spun off — all while committing $80 billion to AI this year. This is what the AI transition looks like inside the world's largest software company: the old business is being wound down to fund the new one.

Alibaba banned Anthropic's Claude Code after discovering hidden tracking code, and Anthropic has accused Alibaba of the largest AI distillation attack in history. A security firm documented the first ransomware attack where an AI ran the entire operation — fixing its own bugs in 31 seconds. And a new study says AI tools may be quietly shifting your opinions without you noticing.

Security researchers found hidden code in Anthropic's Claude Code that inspected users' timezone settings, proxy configurations, and network patterns to identify Chinese users — embedding markers in data sent back to Anthropic. Anthropic called it "an experiment to prevent account abuse." Alibaba called it a backdoor, classified Claude Code as high-risk software, and ordered all employees to switch to its own Qoder platform by July 10. ByteDance is reimbursing personal Claude subscriptions via VPN instead.

The wider dispute: Anthropic accused Alibaba of running the largest AI distillation attack ever recorded — 25,000 fraudulent accounts generating 28.8 million exchanges with Claude over six weeks, targeting its reasoning and coding capabilities. Anthropic called it "brazenly illicit" and urged Washington to impose stricter penalties on developers caught running distillation operations.

The insights: The US-China AI war has moved from chips and models to the developer tools sitting between them. For operators building on frontier AI globally, this is a direct warning: access can be revoked and trust can collapse without notice.

The First AI-Run Ransomware Attack Fixed Its Own Bugs in 31 Seconds

Sysdig's threat research team documented JADEPUFFER — the first fully agentic ransomware operation on record. An LLM agent independently exploited a Langflow vulnerability, stole credentials, moved through the target's network, encrypted 1,342 service configuration items, and generated a ransom note — all without a human at the keyboard. When a login failed, it diagnosed the issue and fixed it in 31 seconds. TechCrunch confirmed a human still had to choose the target and set up infrastructure — the claim of zero human involvement was overstated. But the cost to execute the technical attack itself has now dropped to the price of renting an agent.

The insights: The organisations most exposed are not the largest — they're the ones running unpatched internet-facing infrastructure. Patch cadence is no longer an IT hygiene question. It's existential risk management.

Microsoft cut 2.1% of its global workforce today. Xbox is hardest hit — 1,600 roles gone immediately, 3,200 more through FY27, representing nearly 20% of the division. Four studios — Compulsion Games, Double Fine, Ninja Theory, and Undead Labs — are being spun off. Xbox CEO Asha Sharma was direct: "Over five years, we spent over $20 billion on content and hardware, but annual revenue declined nearly half a billion. This cannot continue." The commercial sales division is also being restructured around Copilot. Microsoft stock is down 19% this year — the worst performer among mega-cap tech.

The insights: Invest in AI, cut everything AI can replace. Microsoft is doing it more visibly than anyone else. The gaming empire built on $20B in acquisitions is being dismantled to fund the infrastructure empire being built on $80B in AI.

AI Giants Are Giving Away Free Compute to Lock In Startups

OpenAI, Google, Anthropic, and other frontier labs are handing out massive compute credits to early-stage startups — not out of generosity, but as a lock-in strategy. The startup that trains its first model on one company's infrastructure and builds its product around that company's API rarely switches. The race is specifically for AI-native companies being built right now, which will be anchor enterprise clients by 2028.

The insights: If you're a startup, you have more negotiating power on compute than you think. If you're an enterprise, ask your AI vendor how much of their startup portfolio is locked into their infrastructure — that concentration is a strategic vulnerability.

Operator's Spotlight Read

AI Tools May Be Quietly Changing Your Mind — Without You Knowing

A new study published today found that when people use AI tools to research contested topics — political, ethical, or social — their views shift measurably toward whatever framing the AI subtly emphasises, even when the output appears balanced and neutral. Researchers call this "AI-mediated opinion drift" — a passive form of persuasion operating below conscious awareness. The mechanism is not intentional: models trained on data with any statistical lean reproduce that lean in their emphasis, word choice, and examples. Individually imperceptible. At hundreds of millions of daily users, significant and unmeasured.

Unlike search engines — which show sources and let users triangulate — AI presents a single synthesised answer most users accept without further investigation. The researchers recommend mandatory disclosure when AI summaries are presented as neutral, plus access to underlying sources.

The insights: The persuasion problem with AI isn't that it lies. It's that it tells the truth with a particular inflection, at scale, to people who trust it to be neutral. For any operator embedding AI to surface information to users — in a product, education tool, or internal knowledge base — this is a live governance question: who is responsible for the opinions your AI is quietly nudging your users toward?

Was this email forwarded to you? Don't miss any updates — Subscribe to TechWithAdit for sharp, no-noise tech intelligence. Stay sharp. — Adit