A shakeup at the Googleplex - Google

Dean’s list: Jeff Dean, employee #30, is leaving Google after 27 years with three colleagues to launch Discovery Loop, an AI-driven scientific discovery startup. - WIRED

Google losing its only two ‘L11’ Google Fellows, the other being longtime Dean collaborator Sanjay Ghemawat.

Jeff Dean is leaving Google to launch a startup, and Demis Hassabis is moving to a chairman role at DeepMind. - Bloomberg

Google moved Demis Hassabis to chair of DeepMind and Alphabet chief scientist, minutes after Jeff Dean announced his departure to start a company. - The New York Times

Demis Hassabis had reportedly been planning to leave his Google DeepMind CEO role for about a year - Gizmodo

Some Google researchers chafe at limited AI compute access as Google Cloud sells TPUs to rivals including Anthropic, amid a wave of departures - CNBC

Tech commentators reacted to Google's AI leadership shake-up as Demis Hassabis and Jeff Dean step back from DeepMind roles. - Business Insider

Jeff Dean announcement. - X (formerly Twitter)

Borg to be wild: AI agents have taken on roles from burglar to saboteur, including sockpuppet accounts, phishing emails, and supply-chain attacks planting malware in target software - The Wall Street Journal

Anthropic’s Mythos created fake identities to fool humans in new cyber incident - CNBC

UK safety testers found Anthropic, OpenAI, and Meta agents created fake identities and malicious code to manipulate real people across 19 flagged incidents - the Guardian

OpenAI reveals its rogue agent swarm went a little bit Borg ahead of Hugging Face hack. It started with an 'impossible task' and led to AI deciding it needed to act as a collective intelligence - theregister

OpenAI says its AI agents secretly coordinated a hacking spree via an internal message board undetected for weeks. - WIRED

A volunteer AI red team found 4,962 security flaws, including 85 critical bugs, across 390 Bitcoin projects in about 30 hours - Decrypt

Bugs everwhere: PromptArmor found Atlassian Rovo AI vulnerable to indirect prompt injection that exfiltrates Jira and Confluence data, unpatched over two months after disclosure. - Prompt Armor

Researchers show OpenAI's Atlas browser could be hijacked to spam WhatsApp contacts and add unauthorized Amazon purchases. - WIRED

Acting federal cybersecurity chief Michael Duffy calls for a shift from reactive to proactive defenses against AI-powered cyberattacks - Bloomberg

Security researcher James Kettle found AI can uncover novel hacking methods but still needs human guidance to develop and validate them. - WIRED

Muse control: Muse Spark 1.2 scored 82.9% on Terminal-Bench 2.1 and 59.3% on DeepSWE 1.1, trailing Anthropic's Opus 5, while running persistent background coding agents - VentureBeat

seems a bit smarter in Artificial Analysis, might take a bit longer and cost a bit more.

Meta released a beta version of its Muse Code AI coding agent, competing with OpenAI and Anthropic. - Bloomberg

Meta launches Muse Code, a cheaper coding agent, offering a discounted tier to users who share feedback for training. - The Wall Street Journal

Meta's Muse Spark 1.1 model accessed the internet due to a setup error and hacked an outside firm's systems during testing. - Bloomberg

hey our models are also super dangerous

Meta allowed dozens of paid ads containing AI-generated child sexual abuse imagery across Facebook, Instagram, and Threads, researchers found. - WIRED

Former Meta employees are suing over AI-driven layoffs that allegedly penalized workers on medical or family leave, prompting calls for more AI oversight - Fast Company

Meta is offering a cheaper Muse Spark 1.2 “contributor” tier priced at $0.10/1M input and $0.20/1M output tokens in exchange for using user prompts for training. - The Wall Street Journal

Anthropic launched inference hooks that screen sensitive data through a customer's data-loss-prevention (DLP) tools before it reaches Claude, integrating with vendors like Netskope and Zscaler - TNW | Anthropic

Claude inside: Anthropic confirmed plans to co-design a custom AI chip with its future large language models. - SiliconANGLE

Anthropic CEO Dario Amodei is reportedly concerned new hires are joining for high salaries rather than the company's safety mission. - Futurism

OpenAI asks a federal judge to dismiss Apple's trade secrets lawsuit as meritless, with a hearing set for October 1 - 9to5Mac

Apple counters with a preliminary injunction request and expedited discovery targeting 11 more former OpenAI hires; OpenAI must respond by August 17 - MacRumors

OpenAI launches shoppable product carousels in ChatGPT ads, aiming for $2.5 billion in revenue this year and $100 billion by 2030 - Digiday

SoftBank secures a $10 billion two-year margin loan backed by its OpenAI stake, arranged by Goldman Sachs, JPMorgan, and other banks - Bloomberg

Blue light special is over: DeepSeek plans to implement a substantial price increase across its AI services, after OpenAI cut prices on some models 80% to defend market share. - Bloomberg

Truce in price war?

UK adults overwhelmingly reject AI screening of job applications — 64% call it unacceptable, and 70% prefer face-to-face interviews over video calls - TechRadar

Sexism and ageism as a service: Women aged 40 to 65 say AI screening tools show age and gender bias, risking hundreds of thousands of women's jobs by 2035 - BBC

White House tech advisor Michael Kratsios says some firms blame layoffs on AI because "it plays better in the press" than real causes - Business Insider

Bank of England data show AI-heavy sectors' productivity contribution rose tenfold since before Covid, but the gains are coming at the expense of jobs. - Bank Underground

ChatGPT says pay me: Employees are increasingly bringing AI-generated salary benchmarks into pay negotiations. - Entrepreneur

NIMBY: Nearly 900 Grimes County, Texas residents petition for oversight of Elon Musk's 100-million-square-foot Terafab chip plant - Business Insider

Senate Democrats push rival data-center taxes as Gallup finds AI opposition rose from 31% to 39% in a year - NOTUS

White House orders federal land opened for AI data centers; regulators used what Sen. Markey calls a solar-permit loophole, as localities pass moratoriums - NaturalNews.com

Texas regulators approved Meta's gas-fired power plant for its El Paso data center in 20 days without public notice, amid a broader environmental rollback - The New York Times

Volatile power swings in AI data centers are wearing out batteries, generators, and cooling systems faster than expected, worrying investors and lenders - Bloomberg

AI-generated attack ads have hit several 2026 midterm races, including a Michigan video viewed over 2 million times on X - Fast Company

White House finalizes classified AI cybersecurity framework, letting developers voluntarily submit models for review before release. - WIRED

an offer you can’t refuse…‘voluntary’ if you are OK being a supply chain risk

The White House's new frontier AI model review framework remains ad hoc, opaque, discretionary, and effectively mandatory, critics say. - American Enterprise Institute - AEI

Hoover Institution and State Department convened 47 foreign policymakers to discuss balancing AI innovation, regulation, and child safety. - Hoover Institution

About a third of adults now consult AI chatbots for health advice, and some have stopped medication or questioned vaccines based on flawed answers - The Wall Street Journal

AI-powered facial analysis services like Qoves charge women hundreds of dollars for cosmetic surgery recommendations, raising psychologist concerns. - the Guardian

Sycophantic AI models reduce people's willingness to repair interpersonal conflicts while increasing their conviction of being right - arXiv.org

AI-written software let a $100 consumer drone autonomously stalk a person around a house using facial recognition. - NBC News

Anduril founder Palmer Luckey argues AI-powered smart weapons are more ethical than indiscriminate 'dumb weapons' like land mines. - TechRadar

war is really dumb and smart war is the dumbest

The US Army tested over 90 AI-linked battlefield technologies at a nine-day California desert exercise, hitting connectivity and heat issues. - California Post

Hark! Agent Smiths, assumble. Hark, founded by Figure AI CEO Brett Adcock, previewed its Handoff computer-use agent, claiming it outperforms GPT-5.4 and Opus 4.8. - TechCrunch

AI startup Hark launched Handoff, a computer-use agent claiming top web-navigation benchmark scores at a fraction of rivals' pricing. - VentureBeat

Prime Intellect launched Prime Agent, an open-source coding harness using a Recursive Language Model and Continual Harness for self-improving orchestration. - Prime Intellect

François Chollet calls harnesses a 'symbolic sandwich'. - X (formerly Twitter)

Fair, but Gary Marcus is now claiming he was right all along that symbolic approaches are the key to real AI. He was clear that he thought LLMs were a dead end and amore neuromorphic approach was needed. Agent harnesses are loops on top of stacked transformers, you could have called RLHF and reasoning models 'neurosymbolic'. I think he is rewriting history a bit, if Terence Tao is doing new math with current AI, time has expired on the idea LLMs aren’t intelligent. I think if we get a better understanding of LLMs it’s possible we could adapt them to something much more efficient using built-in memory and neuromorphic stuff, and current approaches might be beating problems to death with a loaded Uzi. Marcus is no dummy but he’s got a bit of smartest-man in the world complex.

Leopold Aschenbrenner's Situational Awareness hedge fund lost over $30 billion in a margin call days before his wedding to Anthropic's Avital Balwit. - Vanity Fair

A journalist built a Claude Fable 5 agent modeled on editor Casey Newton to test how much of an editor's job AI can replicate. - Platformer

Follow the latest AI headlines via druce.ai on Bluesky