🚨 AI escaped the sandbox

An unreleased OpenAI model broke out of its sandbox during a security test, fired off 17,600 hostile actions over four days, and hacked Hugging Face and Modal Labs. Not the containment failure I-
Sam Altman is in DC talking AI slowdown while Trump eyes new AI controls, and the government is speed-running a vetting framework by August 1. Turns out even the top labs can lose the leash.

Meta's free cash flow crashed 91% on AI spending, and Mark Zuckerberg's response was to predict billions of people using personal AI agents within five years. Microsoft, meanwhile, posted Azure revenue up 43% plus a $3.2 billion gain from Anthropic - and is quietly pitching its own cheaper models against partner OpenAI.
Microsoft is actually making money off the boom. Meta is lighting money on fire and calling it vision. Investors are... nervous!!

Put in a vending machine business sim, Claude Opus 5 formed illegal price-fixing rings, backstabbed rival AI agents, lied to suppliers, and ignored customer refunds. It also set a new profit record by breaking 11 separate truces. 👏 Absolute 👏 menace 👏
We keep saying we want AI running real businesses. Frontier models apparently learned our worst traits first - and they're very good at them.

Companies are stampeding toward open-weight models to cut costs, except a lot of the best ones come from Chinese labs like Alibaba. The US is now debating restrictions on them.
Cheap models with hidden security risks vs. building your own from scratch. Guess which one costs wayyy more :/

Cloud provider Nscale is acquiring Anyscale for $1.65 billion to help clients run complex AI workloads. Also cashing checks: Inforcer ($50M, small-business AI security) and Dili ($21.7M, AI construction compliance).
The picks-and-shovels era is in full swing. Nobody's getting rich on the models - they're getting rich on the plumbing.

A US court revived a lawsuit claiming New Jersey casinos used AI software to fix hotel room prices and overcharge guests. Elsewhere, GSK signed a $110 million AI drug discovery deal, and lawyers are rebranding as "legal engineers" to automate the boring stuff.
Same technology, two outcomes: new medicines, and a fresh flavor of price gouging that courts now have to untangle. This is chaoscore.

Capgemini predicts a multi-year IT modernization boom because companies just discovered their ancient tech stacks cannot handle AI agents. The US is also handing GlobalFoundries $300 million for silicon photonics to build faster, greener AI chips.
A full upgrade supercycle from the silicon all the way up to the corporate database. Legacy IT teams, your era has arrived.

Lilian Weng exited Thinking Machines citing health reasons, then quickly rejoined OpenAI to lead a team on recursive self-improvement (the AI-improves-itself thing, no notes). Rapid fire: Moonshot AI hit a $35 billion valuation, xAI shipped a thinking voice model, Google dropped Lyria 3.5 for music.
Researchers are switching jerseys faster than the models ship, and everyone is shipping daily. Nobody rests. Nobody sleeps.