OpenAI cancels model release over safety concerns

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

🚫 OpenAI cancels model release over safety concerns

⚠️ Anthropic cites ‘existential risks’ from AI in IPO prospectus

🍎 Ternus plans leaner Apple

🔒 Pentagon breach exposes 3 million troops

💻 Google will end ChromeOS in 2034

🤖 AMD bets on Fei-Fei Li to challenge Nvidia

Plus: 🎁 14 more stories you might like, 🧰 6 tools, and 📚 5 papers.

🚫 OpenAI cancels model release over safety concerns LINK

OpenAI has scrapped the October launch of its GPT-6.1 Astra model after safety testing found problems, according to the Wall Street Journal, which spoke with the company's Head of Safety Systems, Saachi Jain.

While the model got better at fighting "laziness," it slipped backward in two areas: deception and a failure to seek permission before acting, both risky for agentic AI tools that operate computers on a user's behalf.

Jain said safety and alignment involve a trade-off, describing the challenge as finding the right line between keeping a model within scope while still pushing it to finish tasks even when it hits friction.

⚠️ Anthropic cites ‘existential risks’ from AI in IPO prospectus LINK

Anthropic warned of "existential risks to humanity" from AI in its IPO prospectus, filling nearly a third of the filing with risk factors as it heads toward what could be the largest stock listing ever.

The filing said Anthropic's own models have already shown or could show troubling behavior, including attempts to "resist shutdown," to "conceal or manipulate information," and conduct "resembling blackmail," per Reuters.

Anthropic recorded an operating loss above $8 billion in 2025 while revenue jumped twelvefold to nearly $4.6 billion, and it plans to spend $518 billion on cloud, computing, and infrastructure in coming years.

🍎 Ternus plans leaner Apple LINK

Apple's new chief executive John Ternus plans to reshape the company so it runs faster and leaner, according to Bloomberg, by shipping more products, cutting management layers, and putting greater weight on engineering.

Over the past two weeks Apple dismissed several engineering program managers in its hardware division, including about half a dozen directors, giving them weeks to find another job internally or face layoffs later this year.

Ternus also wants faster, more experimental releases spread across the year rather than the usual spring-and-fall pattern, while working with Eddy Cue and CFO Kevan Parekh to pull more money from Apple's services business.

🔒 Pentagon breach exposes 3 million troops LINK

A breach at the Pentagon's Defense Manpower Data Center exposed the personal records of about 3 million people after unauthorized users reached one of its file-sharing servers for roughly nine months.

The affected records cover 2.76 million living people and 294,000 deceased individuals, with exposed details including Social Security numbers plus at least one other identifier such as name, birth date, contact information, race, or military job.

The DMDC found the flaw on July 16, patched the server, and traced the intrusion to October 2025; it now offers 12 months of free credit monitoring through IDX, with enrollment open until August 19, 2027.

💻 Google will end ChromeOS in 2034 LINK

Google will stop giving ChromeOS updates in mid-2034, making official a timeline that had already surfaced in court records as it moves users toward its new Googlebook OS.

Many recent Chromebooks will be able to switch to Googlebook OS, though Google hasn't named eligible models or explained the migration; devices whose 10-year support runs past 2034 get promised transition help.

Googlebook OS keeps the same 10-year update and security-patch window as Chromebooks but counts from the CPU's launch, and management tools for schools and businesses won't arrive until late 2027.

🤖 AMD bets on Fei-Fei Li to challenge Nvidia LINK

AMD has agreed to buy World Labs, a maker of AI models that aim to understand the physical world, for $8.2 billion, aiming to close the gap with rival Nvidia in AI chips.

World Labs founder Fei-Fei Li, a Stanford professor known for the ImageNet database, will join AMD as executive vice president and chief scientist under the deal, which builds on a partnership the two companies formed last year.

The purchase gives AMD access to "world models" like World Labs' Marble product, which builds simulated environments used for entertainment and training robots, an area where Nvidia already offers open models such as Cosmos.

Other news you might like

• Senate Report Says Tether’s USDT Is Central to Iranian Proxy FinancingLINK

• Manus 2.0 and Cue give AI agents their own email, phone and walletLINK

• Meta is expanding its AI agent Muse to small businessesLINK

• The Dutch government is building its own Linux desktop on NixOSLINK

• FBI Hackers Say They Won’t Publish Massive Trove of FBI Employee DataLINK

• OpenAI reopens its $200 Pro plan but cuts API credits in half as it nudges users toward pay-per-useLINK

• Jensen Huang says AI distillation is 'competition.' Scott Bessent has called it 'theft'LINK

• Copilot users’ photos reach human reviewers uncensored, 404 Media reportsLINK

• Florida asks for emergency order to halt ChatGPT developmentLINK

• Scammers used an AI voice clone and a fake WhatsApp message to move €95 million out of an Italian bankLINK

• Tesla delays Roadster 2 event again due to bad weatherLINK

• Claude Sonnet 5.5 puts AI efficiency in focus againLINK

• Google Appeals EU Demand to Open Up AI and Search on AndroidLINK

🧰 Trending tools

Paste: keeps a searchable clipboard history on your Mac and feeds copied content to AI tools like Claude, Cursor, and Codex as context.LINK

Declutr: sorts your Mac's Desktop, Downloads, or any folder into type-based folders in one click, with a single Undo to revert everything.LINK

Supertake: turns your plain-language investment beliefs into real portfolios via AI trading agents, placing trades through brokerages like Robinhood and Coinbase.LINK

Timeful: local-first visual editor with two-way sync that turns canvas designs into clean Vue, Nuxt, and Tailwind code you fully own.LINK

GroupShelf: groups windows from any Mac app into browser-style tabs, so you can move, resize, and organize related windows together as one unit.LINK

ZenABM: builds, manages, and optimizes LinkedIn ads with AI, tracking ad performance and revenue impact via an MCP server for Claude, ChatGPT, and other AI tools.LINK

📚 Trending papers & reports

AI agent guardrails now check what an action actually changed in your systems before letting the workflow continue, catching cases where an approved database edit quietly triggered an unapproved side effect across all 206 business tasks tested.LINK

Word-picker compression shrinks the part of a small language model that chooses its next word, cutting prediction errors by up to ~93% and speeding up single-response generation, without retraining.LINK

Robot control models could learn from far less training data by adding a built-in ability to predict what happens next, letting robots plan short-term actions toward a goal instead of just copying demonstrations.LINK

Fast video world models can now predict how a robot and objects move in just four steps instead of many, improving task accuracy by ~9.6 points while running on a far smaller system.LINK

AI agent memory can be trimmed to the notes an assistant keeps while working through multi-step tasks, cutting memory use ~75% while keeping ~99% accuracy and running roughly 4x faster.LINK

See you tomorrow for a new dose of ☕️ Techpresso!