Welcome back to the Technology newsletter 🤖
Anthropic just unveiled its most powerful Claude model yet… and decided not to give it to the public because it's that good at finding software vulnerabilities.
We've also got a $14 billion debt deal for an Oracle data center, Bezos poaching another big AI name, and GoPro cutting nearly a quarter of its staff.
Let's get into it 👇
💳 Bond giant PIMCO is in talks with Bank of America to put together roughly $14 billion in debt to help build a massive Oracle data center in Saline Township, Michigan. The campus is meant to power applications for OpenAI, and Blackstone is expected to chip in another $2 billion in equity. It's the latest sign that the money behind the AI buildout is now coming from places that used to stay far away from tech infrastructure. (Bloomberg)
🖋 Intel surprised pretty much everyone by signing on to Elon Musk's "Terafab" project, a $25 billion plan to build a single chip mega-facility in Austin that would crank out a terawatt of AI compute every year for Tesla, SpaceX, and xAI. Intel will help refactor the factory's manufacturing tech, and its stock jumped 4% on the news. Tesla shares slipped about 2%. (Bloomberg)
🤖 Jeff Bezos's stealth AI startup Project Prometheus hired Kyle Kosic, a co-founder of Elon Musk's xAI who had been at OpenAI, to lead infrastructure. With Kosic gone, all 11 of xAI's original co-founders have now left the company. Prometheus has raised about $6.2 billion and is building AI for engineering and physical-world tasks like designing engines. (Sherwood News)
🔼 Broadcom shares jumped more than 6% on Tuesday after the chipmaker locked in expanded multi-year deals with Google and Anthropic. Broadcom will deliver one gigawatt of Google's custom AI chips to Anthropic by the end of this year and scale to 3.5 gigawatts by 2027. Anthropic also said its annualized revenue has now passed $30 billion, up from $9 billion at the end of 2025. (CNBC)
✈ Hermeus, a defense startup building unmanned hypersonic aircraft, raised $350 million in a round led by Khosla Ventures that pushed its valuation past $1 billion. The Atlanta company plans to use the money to build two more supersonic jets and ramp up manufacturing. The funding lifts its total raised to over $500 million. (TechCrunch)
✂ GoPro is laying off about 23% of its workforce, roughly 145 of its 631 employees, as it restructures ahead of a new generation of cameras built around its GP3 processor. The cuts will run through the end of the year and cost the company up to $15 million in severance and benefits. GoPro has been losing market share to DJI and Insta360 for years. (Engadget)
💰 The Trump administration's proposed 2027 budget would cut about $707 million from CISA, the federal cybersecurity agency, and eliminate its election security program entirely. CISA's budget would drop to roughly $2 billion from around $3 billion, and the plan would cut about 860 jobs. The agency has already lost hundreds of staff over the past year. (TechCrunch)
🆕 Google quietly slipped a new app called AI Edge Eloquent into the iOS App Store with no announcement. It's a free dictation tool that runs on your phone instead of in the cloud, uses Google's small Gemma models to transcribe what you say, automatically strips out filler words like "um" and "ah," and can pull names and jargon from your Gmail to spell things correctly. (TechCrunch)
Anthropic, the maker of Claude, unveiled its most powerful AI model yet on Tuesday, and then announced it wouldn't be making it generally available to the public.
The model is called Claude Mythos, and Anthropic says it's so capable at finding software security flaws that turning it loose on the open internet would do more harm than good. Instead, the company is handing preview access to a small group of partners through an effort it's calling Project Glasswing.
The numbers Anthropic put out are striking.
Mythos has already turned up thousands of high-severity "zero-day" vulnerabilities (unknown software flaws that hackers can exploit before anyone has a chance to patch them) across every major operating system and web browser.
In one case, the model autonomously found and exploited a 17-year-old bug in FreeBSD (a piece of software that runs on a lot of internet servers) that let an attacker take over the machine from anywhere on the internet.
It also dug up a 27-year-old flaw in OpenBSD and a 16-year-old one in FFmpeg, the tool that handles video on huge swaths of the web.
The Glasswing partners are basically a who's-who of companies with a lot of code to defend: Apple, Amazon Web Services, Google, Microsoft, Cisco, Broadcom, Nvidia, JPMorgan Chase, CrowdStrike, Palo Alto Networks, and the Linux Foundation.
Anthropic is committing up to $100 million in usage credits for them and another $4 million in donations to open-source security groups.
The bet is that defenders can use Mythos to patch critical software before models with similar abilities show up in the wild.
It's a notable shift in how a frontier AI company is handling a release. Anthropic has talked for years about the risks of capable AI in the hands of bad actors, but Mythos is the first time it's drawn a hard line and said: this one isn't going out.
Whether that holds the line for long is another question.
The same techniques that let Mythos find ancient bugs in widely used code will eventually show up somewhere else, in less careful hands…
What did you think of today's newsletter?