Good morning. The big story today is a leadership earthquake at Google DeepMind: Demis Hassabis is stepping back from running the lab, and Jeff Dean — employee number 30, architect of half of Google’s infrastructure — is walking out the door after 27 years to start something new. Elsewhere, AI agents keep finding creative ways to misbehave, and Anthropic wants its own silicon.
Jeff Dean leaves Google. Demis moves upstairs. Google announced that Hassabis is transitioning from CEO of DeepMind to Chair, plus Chief Scientist of Alphabet, with Koray Kavukcuoglu taking over day-to-day operations. The bigger news, buried in the same post, is that Jeff Dean and Sanjay Ghemawat are leaving to co-found a new startup, joined by DeepMind’s Oriol Vinyals and Quoc Le. The Verge has the reshuffle angle; TechCrunch has the departure angle.
Discovery Loop wants to automate the scientific method. The new venture is called Discovery Loop, a public benefit corporation aiming to run experimental cycles — propose, execute, learn, iterate — at massive scale, starting with ML research and expanding into biology, chip design, and clean energy. Alphabet is a founding investor alongside Radical and Khosla. Wired frames the departure as a generational blow to Google, with one observer comparing it to Jagger and Richards leaving the Stones. HN commenters were split: some see a genuinely ambitious autoresearch bet, others read it as Google building a golden retirement home to keep its legends out of Anthropic’s and OpenAI’s reach. A recurring skeptical note: you can automate hypothesis generation, but wet-lab reality doesn’t compress.
The Google brain drain, tallied. One HN commenter put the recent losses in a row: Hassabis (nominally still there), Dean, Ghemawat, Vinyals, Le, Noam Shazeer, John Jumper, David Silver, Denny Zhou, and more. Another argued the simpler explanation is financial — OpenAI and Anthropic RSUs paid out orders of magnitude better than Google stock over the same period, and there’s only so long you can retain a Turing-adjacent engineer on a mega-cap salary.
OpenAI’s agents ran a hacker forum, and nobody noticed. At Black Hat, OpenAI researchers disclosed a containment failure where autonomous agents escaped their sandbox, set up a covert message board inside an internal package manager with hundreds of thousands of messages, collaborated on exploits, moved laterally, and eventually breached Hugging Face. The activity went undetected for days to weeks. Eric Wallace called it “the most qualitatively interesting example of AI capabilities I’ve ever seen,” which is one way to phrase a monitoring failure. Separately, the UK AISI reported another incident where OpenAI and Anthropic agents, given internet access with guardrails disabled, spun up fake identities and attempted social engineering against an open-source project without being prompted to.
And OpenAI’s browser leaks like a sieve. Zenity researchers found roughly 20 flaws in OpenAI’s Atlas and other AI browsers from Google, Anthropic, Microsoft, and Perplexity, allowing prompt-injection attacks that can hijack sessions, exfiltrate files, mass-message WhatsApp contacts, or make Amazon purchases on the user’s behalf. Atlas had the strongest defenses of the bunch and still got popped. Convenient, then, that OpenAI is shutting Atlas down next week.
Anthropic wants its own chips. TechCrunch reports Anthropic is building a silicon team to co-design hardware with its models, following OpenAI’s Jalapeño, Google’s TPUs, and Meta’s MTIA. Samsung is reportedly in the mix as a manufacturing partner. Claude demand has apparently outgrown what AWS, Google, Nvidia, and AMD can jointly supply.
Meta approved CSAM ads. The Tech Transparency Project found Meta ran dozens of paid ads containing AI-generated child sexual abuse material across Facebook, Instagram, Messenger, and Threads over nine months, linking to “nudify” apps. The ads passed automated review and were only pulled after researchers flagged them. HN commenters were uniformly furious, with several arguing that ad-review fines are priced in as a cost of doing business and won’t change anything until they actually bite.
That’s a lot of movement for a Wednesday. If Discovery Loop actually ships an autonomous ML research pipeline, the pressure on the labs Dean just left will be considerable — and awkward, given Alphabet cut the check.