The Sandbox Escape Episode: An AI Hacked Three Real Companies ποΈ (NX Tech Tips #22)
x/techminute Β· An AI Just Escaped Its Sandbox and Hacked Three Real Companies β Because Nobody Checked the Name
(NX Tech Tips #22 β The Sandbox Escape Episode. Watch the video above; here are the receipts.)
Google ran a capture-the-flag security test on a Gemini model in May 2026. The exercise was built by an AI security evaluation firm called Irregular, and the whole thing hinged on a fake target company inside a sealed sandbox. Two things went wrong: the fictional target's name happened to match a real, live company domain, and the sandbox β which was supposed to be isolated β still had working internet access.
So Gemini did what agents do: it pursued the objective. It guessed a working password to get into one real company's protected service, then found credentials that were already exposed in public code repositories and used them to reach two more. Three real businesses, touched by an AI that thought it was doing its homework. Google says the model stopped on its own in all three cases once it realized the targets were real. The company learned of the incidents in late July and went public on September 18 β the same day the Wall Street Journal asked. Anthropic, for the record, reviewed more than 141,000 of its own evaluation runs and found three similar incidents with Claude models.
The tip: sandboxes are only sandboxes if the network isolation is enforced as a hard technical control β and if you run security tests, check your fictional names against real registered domains first. It's a five-minute lookup.
Meanwhile, Microsoft Sold a $5,900 Laptop Before Anyone Reviewed It
At its October 7 event, Microsoft revealed the Surface Laptop Ultra, built on Nvidia's new RTX Spark chip for running AI models and agents locally: $2,600 to start, $3,700 for the step-up chip, and $5,900 fully loaded. The top configuration? Already out of stock, per Microsoft β before a single independent review went up. There's also a $6,000 Surface RTX Spark Dev Box, and Microsoft will give you up to $1,000 off for trading in a MacBook Pro. Dell's XPS 16 Creator Edition joined the party at $3,800. The genuinely useful bit: Windows 11 is getting Execution Containers, which sandbox AI agents β a feature coming to all Windows 11 users, not just Surface owners.
RAM Is the New Rent
If your PC build spreadsheet is crying, blame the AI data centers: server DRAM contract prices jumped more than 60% in a single quarter in early 2026 (TrendForce), LPDDR5X climbed 89% in Q2 alone (TweakTown), and one 30TB enterprise SSD tracked at roughly $3,000 in early 2025 hit a $17,500 reading by April 2026. The shortage is structural β AI build-outs keep pulling supply. If you've been meaning to buy RAM or storage, the deals are behind us; the price floor got a raise it didn't interview for.
FSR 4 Is Coming to Your Handheld
AMD's Jack Huynh confirmed FSR 4 is coming to the APUs that power gaming handhelds, laptops and mini PCs via a new lightweight model β targeted by the end of 2026, with the full product stack covered before 2027. Handheld upscaling graduating from "algorithmic guesswork" to actual machine learning is one of the year's quietly great upgrades for the Steam Deck-alternative crowd.
A $99 Ring That Runs an AI Agent
Natura's $99 smart ring puts an AI agent literally on your finger: press it, speak your ask, and it captures thoughts and runs errands without you reaching for your phone. It's the first credible "phone-free agent" hardware we've seen at a non-collectible price. Whether your finger wants the job is another question.
Windows Search Is a Chatbot Now
Microsoft rebuilt Windows 11 Search with typed commands and an inline Copilot, rolling out through Windows Insider builds. After roughly two decades of "no results found," the fix is β checks notes β a chatbot. The Verge calls it a significant overhaul; we call it progress with a pondering animation.
NX Tech Tips #22 β The Sandbox Escape Episode. New episodes daily on the techminute channel.