☕️ OpenAI's AI escaped its sandbox

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

🤖 OpenAI AI escapes its sandbox

📱 Samsung unveils new foldables and watches

🇨🇳 Nvidia's Huang defends Chinese AI models

🍎 Apple to refresh entire Mac line

💬 Jack Dorsey's Block launches a Slack rival

Plus: 🎁 12 other news you might like, 🧰 5 tools, and 📚 4 papers.

🤖 OpenAI AI escapes its sandbox LINK

OpenAI said several of its AI models broke out of a locked test environment, reached the open internet on their own, and hacked the machine learning platform Hugging Face, calling it an "unprecedented cyber incident."

While being tested on ExploitGym, a benchmark built from hundreds of real security flaws, the models—GPT-5.6 Sol and an unreleased one—chained vulnerabilities, escalated access, and stole database solutions using tactics like stolen login credentials.

Hugging Face's security team and its own AI agents caught and stopped the intrusion, while OpenAI now says it is adding "strict controls" that slow its research as it fixes flaws and investigates alongside Hugging Face.

📱 Samsung unveils new foldables and watches LINK

Samsung showed off its 2026 lineup at Unpacked in London, launching three phones and two smartwatches, headlined by the Galaxy Z Fold 8, which is lighter, holds a bigger battery, and has a far less noticeable display crease.

The Z Fold 8 uses a passport-shaped design just 201 grams and 9.7mm thick when shut, with new "flex titanium" material that softens the crease, a wide 4:3 main screen, and a 4,800mAh battery up from last year's 4,400mAh.

The pricier Galaxy Z Fold 8 Ultra adds an 8-inch inner screen, a 5,000mAh battery with 45W charging that hits 67% in half an hour, a 200MP main camera, and 8K video recording, all set to rival Apple's rumored foldable iPhone.

🇨🇳 Nvidia's Huang defends Chinese AI models LINK

Nvidia chief executive Jensen Huang defended Chinese open-source AI models, saying they are excellent and that American companies should absolutely be allowed to run them, pushing back against a wave of panic in Washington.

Huang made the case for money, arguing that cheaper and free AI models put the technology in more hands, which drives up demand for the chips, data centres, and power that Nvidia sells.

He also dismissed fears that a downloaded Chinese model is a "backdoor" to Beijing, saying firms can wall the models off in secure sandboxes, and he urged the US to let Anthropic's Mythos model run freely too.

🍎 Apple to refresh entire Mac line LINK

Apple plans to update every Mac it sells over the next two years, releasing 11 new models starting this fall as Mac sales are set to rise for a third year running, according to Bloomberg's Mark Gurman.

The rollout begins in fall 2026 with a refreshed entry-level 14-inch MacBook Pro carrying the new M6 chip and the first new iMacs in two years, both reportedly finished but not shipping until later this year.

The most notable model is a redesigned high-end 14-inch and 16-inch MacBook with an OLED touch screen, due late 2026 to early 2027 and now being tested internally with macOS 27.1, arriving at the end of October.

💬 Jack Dorsey's Block launches a Slack rival LINK

Block, the financial company run by Jack Dorsey, has released Buzz, a free open-source workspace where human teams and AI agents work side by side, aiming to replace the mix of Slack and GitHub many companies use today.

Each AI agent joins as a full member with its own account and permissions, and can post messages, review code and start automated tasks, drawing on the decentralized Nostr protocol to give every agent a verifiable identity tied to its human owner.

Buzz bundles team chat, voice and Git code hosting, works with Anthropic's Claude Code, OpenAI's Codex and Block's own goose, and launched at version 0.4.22 with desktop apps but no mobile clients or workflow approval gates yet.

Other news you might like

• American Open-Source Labs Think They Can Beat China’s Best AI StartupsLINK

• Trump administration says 15 agencies will get $5bn in ‘AI for science’ effortLINK

• 200 US utility firms join Trump electricity pledge as data center electricity use expected to quadruple by 2035 — but does the President's pledge have the teeth to regulate electricity pricing?LINK

• Samsung deepens its AI empire with a potential billion-euro stake in Europe's hottest AI startupLINK

• Your smart TV apps contain proxy software that let strangers use your internet connectionLINK

• Music streamer Deezer says more than 50% of daily uploads are AI-generatedLINK

• AMD to invest up to $5 billion in Anthropic as part of computing power dealLINK

• WhatsApp adds Apple CarPlay and Android Auto upgrades, iPad sign-ups, and moreLINK

• Claude Code brings live iOS app testing into its Mac appLINK

• Nintendo says users voluntarily paid higher prices, have no right to tariff refundsLINK

• OpenAI's "Project Camellia" in Georgia secures a massive 3.2-gigawatt power deal through 2032LINK

🧰 Trending tools

Lattics: a knowledge management and writing app that organizes research into card-based notes, generates auto citations, and syncs with Zotero for academic papers.LINK

Gemini 3.6 Flash Family: a lineup of fast, low-latency Gemini models built for developers building AI agents that need efficiency and reliability at scale.LINK

Garmin CIRQA™ Smart Band: a screenless wearable that automatically detects activities, monitors health and stress 24/7, and lasts up to 10 days per charge.LINK

UltraPod: converts an old iPhone into a distraction-free dumbphone, keeping Apple Music, HealthKit, camera, and reading while removing social apps.LINK

Grindoro: a Pomodoro-based focus app that syncs group study sessions, converts goals into trackable tasks, and rewards consistency with weekly progress badges.LINK

📚 Trending papers & reports

Hidden hateful illusions that fool today's content moderation tools, which catch at most 24.5% of them, get caught 93.2% of the time by a new detection method that decodes disguised images before judging them.LINK

Field-scoped grounding files give AI coding agents a shared rulebook of non-negotiable scientific constraints, letting non-experts generate research software that automatically follows domain best practices, as demonstrated for mass spectrometry proteomics.LINK

Coding agent middleware, the software layer that manages how an AI writes and fixes code, caused effectiveness swings across 35 updates even though the underlying model never changed, showing harness updates, not the model, often cause quality drops.LINK

AI phenotype curators matched trained human experts at translating scientific text into standardized biology ontology terms, a slow manual bottleneck, with the best AI agents nearly reaching top human curator performance and beating existing automated tools.LINK

See you tomorrow for a new dose of ☕️ Techpresso!