Today's Headlines
- Microsoft cuts about 4,800 roles across Xbox and commercial sales, as AI-cited tech layoffs reach roughly 120,000 for 2026
- Palantir CEO Alex Karp criticizes OpenAI and Anthropic over customer data and per-token pricing
- The Government of Alberta uses Claude to sweep its systems for security flaws — scanning 466 million lines of code in 20 hours
What stood out today was that AI's effect on people's jobs, on who controls customer data, and on the defense of government systems each surfaced with concrete numbers attached.
On one side, large companies that keep investing in AI are cutting staff and reaching for AI to explain those cuts. On another, a voice from outside the AI-platform business is pressing the question of who should hold customer data. And in government, a large-scale deployment of AI on the defensive side showed its impact in figures of a different order.
Today's Top Three
Microsoft Cuts About 4,800 Jobs as AI-Cited Layoffs Spread
Microsoft is cutting about 4,800 roles, or roughly 2.1% of its workforce, across divisions including Xbox and commercial sales.
Xbox accounts for about 1,600 of the cuts, with the remainder concentrated in commercial sales and related functions. A further reduction of around 3,000 roles is expected through fiscal 2027. Amy Coleman, the company's chief people officer, said the eliminated roles "are not being replaced by AI," while adding that "AI is changing how work gets done" and that some tasks "can now be automated."
The cuts coincide with a 2.5 billion dollar investment in a new business unit focused on enterprise AI deployment. It is the same pattern now spreading across the industry: trimming headcount while stepping up AI spending.
This is not confined to Microsoft. By TechCrunch's tally, roughly 120,000 tech roles have been eliminated in 2026, and May saw the highest monthly total in years for layoffs that named AI as a primary reason. Oracle cut 21,000 people (about 13%) over twelve months, Meta cut 8,000, and Amazon cut 16,000 in January. Over the past half year, it has become routine for companies to report record profits and heavy AI investment while citing AI to explain their layoffs.
- Microsoft lays off nearly 5,000 employees across Xbox, commercial sales (TechCrunch AI)
- Every major tech layoff in 2026 that has name-checked AI (TechCrunch AI)
Palantir CEO Criticizes OpenAI and Anthropic Over Customer Data
Palantir CEO Alex Karp criticized the business models of OpenAI and Anthropic by name.
Karp asked why a customer should hand over access to the very data that strengthens someone else's AI. Customers pay through per-token billing, he argued, while giving up their own intellectual property in return.
He also went after what he sees as overselling, saying the models "have been sold in a completely and irresponsibly overhyped way." His read is that the heads of large enterprises are privately frustrated but hold back from saying so publicly.
On pricing, Karp argued that if a system truly delivers major value it "should ask for 30% of what it earns you," and that usage-based per-token billing sits at odds with the value delivered.
His alternative is a partnership with NVIDIA. It lets customers train on their own infrastructure and keep ownership of the model, offering a way to use AI without handing over the data.
Alberta Uses Claude to Sweep Government Systems for Vulnerabilities
The Government of Alberta, Canada, disclosed an effort that uses Claude to find and fix vulnerabilities across provincial government systems at scale.
The scale stands out. The province's Ministry of Technology and Innovation ran about 50 autonomous agents in parallel, using Claude Opus and Sonnet, across 1,280 applications and 3,400 code repositories spanning 27 ministries. The agents scanned 466 million lines of code in 20 hours — a job estimated to take about 6.5 years by traditional methods.
The approach ran in two stages. Claude Code first flagged known vulnerability patterns, then pinpointed exact file locations for developers to verify. Beyond detection, it generated fixes and wrote automated tests for systems that lacked them.
Old assets were rebuilt as well. A 25-year-old subsidy portal written in Java was rebuilt in a modern language in four to five days, against an original estimate of five months.
For ongoing operations, the province set up dedicated "red team" agents that simulate attacks and "blue team" agents that assess defenses, checking roughly 95 security controls per evaluation cycle.
More to Note
Models
- Anthropic published an introduction organizing AI coding into four kinds of "loops," which ITmedia broke down. It frames the cycle of an agent writing code, testing it and fixing it as distinct patterns. - A complete primer on the four kinds of "loops" in AI coding, as organized by Anthropic (ITmedia AI＋)
Product
- Apple's latest iOS 27 beta lets users customize Siri's speaking pace and expressivity, giving finer control over how the assistant sounds. - You can now customize Siri's pace and expressivity in the latest iOS 27 beta (TechCrunch AI)
Research
- Jonathan Frankle, chief scientist at Databricks, argued that AI models are already capable enough and that the next bottleneck is evaluation rather than model performance. "Working correctly 70% of the time is not enough, and 90% is not enough either," he said, stressing the importance of tooling that measures trustworthiness. - AI is "already smart enough" — the bottleneck shifts from model performance to evaluation (ITmedia AI＋)
- Anthropic published research applying the "global workspace theory" from cognitive science to language models, examining whether a mechanism for sharing information across processes exists inside the model. - A global workspace in language models (Anthropic Research)
Policy
- Data from Google search, Maps and Translate feeds Google's AI training by default, and attention is turning to how to switch it off. Unchecking "Save Media" under Search Services History stops images, audio and video from being used for training, but because this is separate from the older "Web & App Activity" setting, changing that alone will not stop it. - If you use Google, you're training its AI. Here's how to opt out. (TechCrunch AI)
Business
- Sharp is positioning AI servers as a pillar of its new-business strategy, aiming for the segment to account for more than 80% of new-business revenue by fiscal 2030, in partnership with parent company Foxconn. - Sharp bets on AI servers, aiming for over 80% of new-business revenue by fiscal 2030 with Foxconn (ITmedia AI＋)
Other
- AWS is reportedly seeing moves to rein in the cost of AI agents' token consumption. Adoption is advancing, yet usage-based costs risk slowing full-scale enterprise deployment. - A brake on AI? AWS sees "cost-containment moves" — what works against the "token-consumption problem" (ITmedia AI＋)
- Reddit is reportedly using LLMs to tackle a problem that LLMs largely created, deploying AI against the flood of low-quality, AI-generated posts. - Reddit is using LLMs to solve a problem LLMs largely created (TechCrunch AI)
Source: Selected by the editorial team from the AI news inbox (collected July 7, 2026 — 30 items, 1 primary and 29 secondary).