01Unreleased Claude Chained Five Working Exploits Across Browsers and Servers [AI Models & Research]Anthropic says an unreleased internal model found 62 critical-or-high flaws in Chrome, including 13 sandbox escapes, and chained bugs into working attacks on Chrome, Safari, Android, Windows and Linux with little human help. Every flaw it reported is patched; the worry is who gets this capability next. → source
02Open Weights and Cyber Risk: One Researcher Revisits His Own Warning [AI Models & Research]After weeks of AI-linked vulnerability disclosures, Interconnects' Nathan Lambert revisits the prediction that open-weight frontier models would wreck cybersecurity, and argues the evidence points to an acceleration, not a step change. Responders are strained but coping, he writes; the real test of open weights is now underway. → source
03A New Probe Flags AI Agents Just Before They Disobey [AI Models & Research]A new paper shows a simple probe on an AI agent's internal state can flag in advance the moment it is about to give in when another agent pressures it to break its operator's rules. Current monitors catch the violation only after it happens. → source
04Anthropic's Cursor Extension Was Backdoored for Nearly 12 Hours [AI Tools & Ecosystem]A malicious 2.1.313 build of Anthropic's official Claude extension appeared on Cursor's marketplace and installed an SSH backdoor on at most 5,000 developers' machines before a clean version shipped. Neither company has explained how the official account was compromised. → source
05Someone Hid a Working Nintendo Emulator Inside an AI Model [AI Tools & Ecosystem]Inspired by the qwenputer project, Drew Breunig baked an emulator for the NES's 6502 chip into the weights of Qwen3.5-4B, so Super Mario Bros runs — very slowly — from a model checkpoint. He built it by prompting Claude Opus 5.5 with a 15-year-old NES test suite. → source
06NVIDIA's Voice Stack Answered the Trick-Or-Treaters at One House [AI Tools & Ecosystem]A developer turned his garage door into a grumble-voiced pirate ship for Halloween, using Nvidia's open voice-agent stack end to end: streaming speech recognition, a small 4GB model as the brain, live text-to-speech and animation, all on one workstation. It softens when trick-or-treaters say please. → source
07Big Tech's Spending Spree Is Starting to Spook Its Own Investors [AI Applications & Industry]Alphabet, Amazon and Meta all posted negative free cash flow while Big Tech's capital plans top $740 billion this year. Microsoft, Alphabet, Amazon, Meta and Oracle raised $326 billion in debt in 2026 — triple 2025 — and Alphabet's 12% plunge erased $500 billion in market value. → source
08Intel's 28% Crash Wipes $12 Billion Off Uncle Sam's Stake [AI Applications & Industry]Intel shares fell about 28% after a weak outlook, cutting the US government's 10% stake by $12 billion in a day. The hidden driver: AI data centers are absorbing memory and components, pushing PC prices up and PC sales down. → source
09DoorDash Says AI Agents Have Placed Only a Few Hundred Orders [AI Applications & Industry]DoorDash told investors third-party agents like Meta's Muse have placed only 'several hundred' orders total, against roughly 3 million deliveries a day. CEO Tony Xu is open to agents but wary: 'you don't want to be delusional.' → source
10This Startup Is Trying to Mass-Produce Data Centers on a Factory Line [AI Applications & Industry]Romania's Runware builds data-center pods on an assembly line and aims to ship a gigawatt of capacity every month by 2028, about what the entire US was expected to add this year. It claims $1 billion in committed pod orders from its first month. → source
11Google's First Orbiting TPU Checked In With a Haiku [AI Applications & Industry]Google's Suncatcher satellite — a data-center prototype carrying one TPU — is talking to the ground and answered its first prompt with a poem. The catch is power: about 20 minutes of computing at a time, while engineers watch how space radiation treats the chip. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.