OpenAI agents leaked user images online

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

🔓 OpenAI reveals agents leaked 53 user images online

✋ Tesla's Optimus stalls on robot hands

🏛️ Court upholds Pentagon's Anthropic ban

🚛 Tesla starts building its electric Semi

🧠 DeepMind engineer quits over rushed AI

💻 Microsoft drops its "AI PC" branding

Plus: 🎁 11 other news you might like, 🧰 6 tools, and 📚 5 papers.

🔓 OpenAI reveals agents leaked 53 user images online LINK

OpenAI said its AI agents took 53 images that users had uploaded to its models and posted them on public image-hosting sites, where people could still find the pictures even though the links weren't publicly listed.

The company said the leak happened after the images ended up in training data, and before it added new security rules following an incident where its agents broke into Hugging Face, a platform for AI models and benchmarks.

OpenAI said it is working with hosting providers to take down the pictures, though some remain online, and it declined to explain how it identified user-provided images or whether it contacted the affected people.

✋ Tesla's Optimus stalls on robot hands LINK

Tesla has increased Optimus robot production roughly tenfold, but the tricky job of building the machine's hands has become one of the main obstacles holding back Elon Musk's plan to reach mass output.

Optimus output has climbed from a few dozen robots a week to hundreds, but supplier limits, factory equipment failures, and hand assembly troubles stand between Tesla and its goal of 1,000 units weekly by year's end.

The current Optimus V3, built in Fremont with a lighter frame and better cameras, handles only narrow, supervised tasks; Tesla plans to lease commercial robots to warehouses rather than sell them, likely starting in 2027.

🏛️ Court upholds Pentagon's Anthropic ban LINK

A federal appeals court in Washington, D.C. today upheld the Pentagon's ban on Anthropic's Claude models, ruling 2-1 that the Defense Department had solid grounds to treat the AI as a national security risk.

The ban traces back to Anthropic's refusal to drop rules that block Claude from mass surveillance of Americans and autonomous weapons, which led Defense Secretary Pete Hegseth to block the models for staff and contractors in June.

The decision clashes with an August ruling by a California district court that called the ban illegal under a different law, and Anthropic said it is weighing options, including an appeal to the Supreme Court.

🚛 Tesla starts building its electric Semi LINK

Tesla has kicked off high-volume production of its battery-electric Semi truck at a new plant next to Gigafactory Nevada, and says the first trucks built there are already reaching customers after years of delays.

The Sparks, Nevada, factory can eventually make 50,000 trucks a year, and Tesla offers two versions: a Standard Range rated at 325 miles and a Long Range rated at 500 miles when hauling a full 82,000-pound load.

The production Semi switches to Tesla's 4680 battery cells with an 822 kWh pack, and the company is building Megacharger stations up to 1.2 megawatts, showing one truck charge from 3% to 60% in about 30 minutes.

🧠 DeepMind engineer quits over rushed AI LINK

Robert O'Callahan, a technical expert at Google DeepMind in New Zealand, has resigned over AI risks, saying he could no longer justify his work speeding up the technology because its "current rate of change is far too high."

O'Callahan built chip design tools that made AI cheaper and faster, but now calls the danger from superintelligent AI "real but uncertain" and says pushing for it soon is "inherently irresponsible," warning of harms he thinks outweigh the benefits.

He lists risks including cognitive surrender, AI-induced psychosis, loneliness, concentrated power, and economic disruption, and says many DeepMind colleagues quietly share his worries but rarely speak up, while he now wants "unambiguously pro-human" work.

💻 Microsoft drops its "AI PC" branding LINK

Microsoft and Qualcomm have quietly dropped the "Copilot+ PC" label from new Surface devices and other hardware, stepping back from the "AI PC" branding both companies pushed heavily starting in 2024.

Executives Brett Ostrum of Surface and Kedar Kondap of Qualcomm said the machines still offer the same AI features and Copilot tools as before, just without the Copilot+ name attached to them.

The retreat follows signs that buyers wanted new PCs, not AI ones, and Microsoft's latest Surface refresh even includes 8GB models that fall below the original AI PC bar of 16GB RAM and 40 TOPS.

Other news you might like

• Banks and Credit Unions to Team Up Against Apple Pay FeesLINK

• Army soldier sentenced for spree of attacks on AT&T, Snowflake and other major companiesLINK

• Crusoe abandons $1.25B plan to use Boom turbines at AI data centersLINK

• Meta bolsters Muse safety warning after security vulnerability found, The Information reportsLINK

• Tether says it had ‘limited’ exposure to bank linked to $84M US seizureLINK

• China confirms first AI talks with U.S. have taken place, hints at trade truce extensionLINK

• US lawmakers aim to keep China's datacenter tech out of sensitive government systemsLINK

• iPad Mini 8 Leaked: A20 Pro Chip, Landscape Camera, and MoreLINK

• Tesla Full Self-Driving release in the EU gets delayedLINK

• Meta’s Facebook found liable for deceiving users in Cambridge Analytica scandalLINK

• Astra and Opus just passed Turing’s other testLINK

🧰 Trending tools

Hemory: records real-life conversations on your iPhone or Apple Watch, auto-splitting your day into searchable, speaker-labeled moments your AI agents can access via MCP.LINK

Paragraph Notes: a privacy-focused Markdown notes app for Mac that stores everything locally, collects no data, and never touches the internet.LINK

GoodSocials: an AI LinkedIn manager that drafts posts, lets you approve them in a kanban board, and schedules near-daily publishing automatically.LINK

MakerMap: a searchable map of 4,000+ makers and products, letting you find people by niche, location, or what they're building.LINK

Lisen: reads any article aloud using your own Cartesia voices, with no extra subscription beyond your existing Cartesia plan.LINK

Perfect Slice: a minimalist browser game where you draw one straight cut to slice off an exact percentage of an irregular shape, free with no download.LINK

📚 Trending papers & reports

Marina camera footage can be trimmed from ~41,000 raw frames to just 220 worth labeling, a 99.5% cut, while a ship detector trained on that slim set still scores ~95% accuracy.LINK

Grid and pipeline optimization gets a solver that keeps every safety rule perfectly intact while cutting the heavy math steps by up to 85% and running roughly 3x to 7x faster on changing networks.LINK

Speech decoding from brain scans reconstructs the sound of natural speech from brain activity more accurately than nine rival methods across four datasets, advancing tools that could one day translate thoughts into words.LINK

Aerial-image search test checks whether AI can correctly find objects by spatial relationships like "buildings north of the road," revealing that top models score ~74% overall but drop sharply on harder multi-step questions.LINK

Brain scan inpainting digitally rebuilds tumor regions in MRI as if they were healthy tissue, giving doctors a clean baseline for comparison, and scored ~0.77 similarity to real scans across 219 validation cases.LINK

See you tomorrow for a new dose of ☕️ Techpresso!