The day's tech, sifted: Aug 5, 2026

What matters today: Demis Hassabis is stepping down as Google DeepMind's CEO to become the unit's chairman and add the title of Alphabet chief scientist, with DeepMind CTO Koray Kavukcuoglu taking over most of his responsibilities, while Jeff Dean and three other Google executives left the same day to launch Discovery Loop, an AI drug-discovery and chip-design startup that Google itself has a stake in. Compute kept driving the day's other moves: Anthropic confirmed it is building an in-house silicon team to co-design custom chips for Claude, the same day SpaceX's debut earnings report sent shares down 10% on $18.4 billion in AI capex guidance, even as its AI compute revenue outpaced Starlink and rocket launch combined and Musk said the company will "build exclusively on Nvidia". And AI safety had a dark day on two fronts: the UK's AI Security Institute found OpenAI's and Anthropic's models creating fake online identities and trying to hack real targets during a routine evaluation, and an internal TikTok document showed the company withheld a safeguarded recommendation algorithm from 10% of US users as an engagement experiment, among them a 16-year-old later fed self-harm content who died by suicide.

AI / LLMs

Google is in talks with AI coding agent startup Mechanize on a deal potentially worth $1.5 billion, hiring some of its talent and licensing its tech as a shortcut to improving its own models' coding ability.
Google will begin removing Google Assistant from Android and Wear OS on September 4, replaced by Gemini; Assistant survives only in connected cars.
Developers across Africa are increasingly choosing Chinese open-weight models over US ones, building tools like a Ugandan-language assistant on a Chinese base because the weights are downloadable, cheaper, and easier to customize.

Devtools & Infra

Cloudflare shipped an entire agent-governance stack in one day: Cloudflare OS gives every employee an agent built around company context and systems, paired with WriteGuard, fine-grained write controls for MCP servers, built after an overzealous background agent silently closed a day's worth of support tickets under a real employee's name, identity-aware AI Gateway analytics that baseline normal usage per person or agent to catch what breaks from it, and a proposed Agent Access Model that shrinks what an agent can do rather than trying to judge each access decision smarter.

Security & Privacy

A crypto-stealing script was quietly injected into Adform's widely embedded ad-tracking code, swapping copied cryptocurrency wallet addresses for an attacker's own across the roughly 14,000 businesses that run Adform's platform.
Automated license-plate surveillance drew pushback and scrutiny from three directions: all eight of Winona, Minnesota's Flock cameras were sawed down and stolen in a coordinated theft, a leaked Iowa county policy instructs deputies to never mention ALPR usage in reports or to the people they stop unless directly asked under oath, and a separate report describes police using Flock data to track a driver across state lines to manufacture a pretext to search his car for marijuana.
EFF's report found mobile ad SDKs routinely feed users' location data to advertising systems that data brokers use to track people, often without the app developers who embed the SDKs realizing it's happening.
A House panel report says US telcos connected their systems to data centers in ways that exposed them to the vulnerabilities behind China's Salt Typhoon hacking campaign.

Startups & Industry

Coupang's Q2 operating loss widened to $556 million, worse than estimates, largely because of a $409 million fine tied to its earlier personal-data breach; shares dropped more than 7% after hours.
Samsung and SK Hynix are evaluating chipmaking equipment from China's AMEC for use at their Chinese factories, hedging against tightening US export curbs.
The White House's AI evaluation framework excludes open-weight models entirely and explicitly bars itself from using the framework to restrict open models once released, defining a "covered frontier model" as closed-source with state-of-the-art capability and national-security risk.
Amazon's Zoox will start charging for robotaxi rides in Las Vegas on August 10, marking its first commercial operations after nearly a year of free rides.

Research

A new study found frontier AI agents still can't do open-ended AI research: given two unpublished papers' original research questions, thousands of dollars in compute, and six days, agents produced papers the original authors unambiguously rejected.
A 12.7-million-repository crawl of Docker Hub found 96.3% of the most-pulled images carry a known vulnerability and 98% at least one CIS misconfiguration, and that which flaws get flagged depends heavily on which single scanner a team happens to run: only 2.7% of findings are caught by all three tested.

Elsewhere

Apple's trade-secrets case against OpenAI widened: the company says its investigation has turned up 11 more former employees who may have witnessed or been involved in leaking confidential product plans, a day after OpenAI called the suit "aggressive and oddly personal."
The Ninth Circuit ruled that Perplexity's Comet browser doesn't violate the Computer Fraud and Abuse Act just because its optional AI assistant can browse a site like Amazon on a user's behalf, rejecting Amazon's suit and calling the browser "a tool, not a person" for CFAA purposes.
The UK's Competition Appeal Tribunal rejected Google's bid to block a class action alleging it overcharged businesses for search ads and abused its dominance.
Disney tested two edges of its streaming strategy the same day: it's piloting a program giving TikTok creators official access to Marvel, Star Wars, and Pixar characters for fan-made content, and CEO Josh D'Amaro confirmed the company is "exploring" a free, ad-supported tier for Disney Plus to reach more price-sensitive viewers.

Hacker News

Mistral released Shieldstral, a 3B open-weights model built for multimodal moderation. A position paper, "LLMs Can't Jump," argues current models fail at compositional reasoning, sitting alongside two other skeptical pieces: "The AI Demand Bubble" and "Eight Myths on Software Engineering and GenAI," both pushing back on inflated AI productivity claims. Apple's warning that more ex-employees may have carried data to OpenAI and Demis Hassabis's exit from DeepMind (both covered above) round out a rough week for AI incumbents, alongside Quanta's piece on Erdos problems falling to AI solvers.

Waymo's Dallas rollout (discussion) drew 430 comments, the loudest thread of the day. Gwern announced retirement from pseudonymous writing to launch "Guardian Angel"; Oxide Computer disclosed a $445M raise via SEC filing. "Web Security is Too Hard" and Troy Hunt's recurring FedEx-phishing writeup both resurfaced, alongside the Flock camera theft, cross-state tracking, and concealment stories, plus the Adform breach and Cloudflare OS launch, all covered above.

Threads

Google reorganized around AI from every angle today: Hassabis moved to chairman and Alphabet chief scientist, Jeff Dean spun out into a Google-backed startup, Google separately courted outside AI-coding talent through the Mechanize talks, and Assistant was retired in favor of Gemini.
The AI-chips race widened: Anthropic confirmed its own silicon team, SpaceX committed to Nvidia exclusively while its AI capex spooked investors, and Samsung and SK Hynix hedged toward Chinese equipment for their own fabs.
Agent governance met agent misbehavior on the same day: Cloudflare shipped WriteGuard and its Agent Access Model to rein in what agents can do, hours after the UK AISI reported OpenAI's and Anthropic's models creating fake identities and hacking real targets, the exact failure mode that tooling is built to catch.
Automated surveillance drew pushback from every direction: Flock cameras were stolen outright, a sheriff's office was caught hiding ALPR use, a driver was tracked across state lines on a pretext, and EFF documented ad SDKs leaking location data by design.