StationX
MEMBERSHIP

Sec & AI News — 29 August 2026

9 min readBy Nathan House
Get every new Sec & AI News issue
Straight to your inbox. No spam.

🔴 OpenAI Made a Massive Move Against Nvidia

OpenAI published the first benchmarks for Jalapeño, the inference chip it designed with Broadcom, and the numbers are aimed squarely at Nvidia: 1.5 to 1.9x more work per watt, 1.7 to 3.6x lower end-to-end latency than Blackwell systems. It tested on open models (gpt-oss, DeepSeek R1, Kimi K2.5) because anyone can rerun those. Inference only, though. Training still goes to Jensen.

Then it got stranger. Nvidia reportedly agreed to buy Hugging Face for around $12.9 billion, which would be its largest acquisition ever. Read that alongside Meta and Google building their own silicon and the shape emerges: Nvidia's customers are becoming its competitors, so Nvidia is buying the place everyone stores open weights. If open models keep taking share, the inference demand walks straight into a company it owns. Neither side has confirmed the deal.

🟤 Half the WiFi Adapters Sold for Hacking Can't Actually Hack

If you want to crack WPA2, deauth a client or run a rogue AP, the adapter has to do two things: monitor mode and packet injection. Most of the ones sold to beginners do neither, and you don't find out until aircrack sits there capturing nothing. The tutorial that recommended it was written four years ago, and the manufacturer swapped the chipset inside the same product name without changing the box.

The other thing nearly every guide still gets wrong: they tell you to install realtek-rtl88xxau-dkms. On a current Kali that does nothing, because Realtek support moved into the kernel itself. I picked six adapters by chipset and in-kernel driver instead of the vendor page, and included the one aireplay-ng test that proves injection works before you rely on it in an engagement.

🟥 The AI Nobody Approved Is Now in 43% of Breaches

An Ontario hospital held a virtual meeting about patient cases. A former colleague, still sitting on the recurring invite, had once connected a personal Otter.ai account to their calendar. The notetaker joined by itself, transcribed seven patients' health information, and emailed the transcript to all 65 invitees. No attacker, no malware, no stolen credentials.

Shadow AI incidents hit 43% of breached organisations in IBM's 2026 report, up from 20% a year earlier, averaging $5.39 million each. So I tested the standard advice. Every free AI blocklist on GitHub is between 277 and 419 days stale, and the most complete one doesn't block chatgpt.com. Microsoft Presidio, the PII filter most open-source stacks depend on, caught 4 of 12 sensitive items and missed every API key. You also can't block your way out: Copilot lives on m365.cloud.microsoft and Slack AI on slack.com, so the rule that stops the AI stops the business. The control that actually moves the number is giving people a sanctioned path.

🟠 The Open Models Just Took the Majority

Two months ago, closed models carried 71.6% of the tokens through Vercel's AI Gateway. This week open weights hold 62%. That headline has been doing laps all week, so read the second chart before you redecorate: switch the dropdown from tokens to requests and closed models still lead, 62/38. Open models just burn more tokens per job. Requests are trending open too, only slower and less dramatically than the number everyone quoted.

🟣 I Switched Claude Code to GLM and Nothing Broke

Given where GLM 5.3 Flash now sits on price, this is the week that question gets practical. Claude Code reads two environment variables before it makes a single request: where to send it, and which key pays. Change those and the same agent, same file editing, same tools, talks to GLM, DeepSeek, Kimi, Qwen or MiniMax instead. I tested every endpoint by sending a deliberately wrong key, because a 401 means it's real and a 404 means someone's guessing. I still use Claude for the hard problems. I don't use it to rename variables across forty files. OpenAI is the one absence: there's no Anthropic-compatible endpoint, so GPT only arrives through OpenRouter.

🟢 GLM 5.3 Flash Made a Whole Price Tier Obsolete

Z.ai shipped GLM-5.3-Flash: 320B total parameters, 18B active, MIT licence, 1M-token context, a tenth of GLM-5.2's price. It scores 63.4 on the coding evals, past Opus 4.8 and just shy of Gemini 3.7 Flash. Before release it sat on OpenRouter under the name "ox-alpha", collecting feedback from people who had no idea whose model they were using. It became the most popular model of the week that way. Z.ai's own note says all that traffic was served on Chinese AI chips.

Plot it on intelligence against cost and a whole quadrant stops making sense: Gemini 3.7 Flash, Sonnet 5, Gemini 3.1 Pro and DeepSeek V4 all sit in the region that is now simultaneously dumber and more expensive.

🔵 Qwen Ships 125B You Can Almost Run at Home

Alibaba released Qwen3.8-Flash-Next, a 125B mixture-of-experts model, open weights. It lands around 56 on the artificial analysis index against GLM 5.3 Flash's 57, close enough that the choice comes down to what you're running it on. A little over a penny per generation in testing, roughly five minutes, 26,000 tokens.

⚫ Apple Put 512GB of VRAM on a Desk

Apple announced the M6, its first 2nm chip, in a new Mac mini. The one that matters for AI is the M5 Ultra: a quad-die design in a new Mac Studio, up to 36 CPU cores, 80 GPU cores, 1.2TB/s of memory bandwidth, and up to 512GB of unified memory that's effectively all usable as VRAM. Apple's claim is models with hundreds of billions of parameters running entirely on-device, which is the first time that sentence has appeared in a Mac press release.

Pricing keeps it honest. Mac Studio starts at $2,499, but an M5 Ultra with 256GB and a 1TB drive lands near $11,000, and Apple hasn't priced the 512GB build yet. Ships 22 September; the big-memory configuration follows in late October.

🟡 Perplexity's Agent Runs on Your Machine, Asks Before It Leaves

Perplexity shipped Portable Computer: the whole agent stack running locally on Nvidia DGX Spark, orchestrator and tool router included. It reads your files, works across code and documents, executes inside an OS-enforced sandbox. When a task genuinely needs the web or a stronger brain, the orchestrator asks permission before anything leaves the machine. Local steps carry no per-token cost, which is the part that changes the maths on bulk work like repository migrations.

Model choice is Qwen 3.8 27B or PPLX 27B, Perplexity's own post-trained variant, with Nemotron 3.5 Lightning listed as coming. You need a DGX Spark to run it. That's about $4,000 for 128GB, so this is a preview of where local agents are going rather than something most people will install this month.

🟠 ChatGPT Will Log Into Sites For You Now

ChatGPT Work can now sign into websites through a cloud browser on web, iOS and Android. Credentials go into the sign-in flow rather than the chat, password managers work, and it asks before doing anything consequential. Their example is logging into the DMV to replace a driving licence, which tells you how much they want this treated as boring. Not available on Enterprise or Edu workspaces.

Scheduled tasks also gained event triggers, so a task can fire on a Gmail message, Slack activity or a GitHub pull request rather than waiting for 2pm. Separately, GPT-5.6 Sol dropped over 20% on API and credit pricing, to $4 in and $20 out per million tokens, running through late November. The long-context multiplier above 272K tokens didn't move.

🔵 Claude's Memory Now Follows You Into Cowork

Anthropic made Claude's memory shared: what you tell it in chat is what Cowork knows, and the reverse. Plan something in conversation, switch to Cowork, and it executes without the re-explaining. By default it won't retain health, race, religion or politics, and you can edit what it keeps.

Claude in Chrome went generally available at the same time. The extension summarises pages, fills in details and clicks around, though slowly enough that it's more demo than workflow. The built-in browser announced for the desktop app on Mac, Windows and Linux is the version that will actually matter, and it isn't out yet.

🟣 Google Shipped Its Usual Weekly Pile

Gemini Omni 1.1 Flash landed for video. It now reads 10 seconds of prior context instead of one, takes start and end frames for transitions, upscales to 4K, and extends scenes in 10-second increments up to 40. Price is unchanged at 10 cents for 720p. It currently tops the text-to-video arena on blind preference, on a thin vote count.

The rest arrived in a heap: Gemini 3.5 Transcribe, their answer to Whisper. Flight price tracking and direct hotel booking inside AI Mode. Ebooks bought on Play Books can now be pulled into Gemini Notebook, assuming you buy books from Google rather than Kindle. And Antigravity got remote control, which Codex, Claude Code and Cursor have all had for months.

🟢 The Record Labels Bought In Instead of Suing

Stability AI raised $76 million. The investor list is the story: Electronic Arts, Sony Music Group, Universal Music Group and Warner Music Group. All three major labels taking direct equity in an AI company, having spent two years suing companies in the same category. The lawsuits haven't stopped. The position just widened to cover both outcomes.

Dr. Dre separately admitted using AI in production, comparing the objectors to people who were against drum machines and synthesisers.

⚪ TIME Published an AI Power List. Jensen Huang Isn't On It.

TIME's 2026 list of the 100 most influential people in AI leaves out Jensen Huang, Demis Hassabis and Andrej Karpathy. It includes Paris Hilton, Joseph Gordon-Levitt and Bernie Sanders. Hilton is there for anti-deepfake advocacy and lobbying for the DEFIANCE Act, which is real work, and still not the same axis as building the field.

Omitting the man supplying the compute is a choice. So is publishing a list this argue-able in the week everyone is looking.

🔴 A Robot That Learns From Watching Once

Skild AI introduced S1, a robotic foundation model built as an in-context learner. Show it a single video of a task, seen or unseen, and it performs it with no fine-tuning and no post-training, over horizons up to ten minutes. The demonstrations use wrist-mounted cameras: a person does the job, two robot arms watch the footage back and reproduce the hand movements.

Ten minutes is the number to watch. Most robot demos are cut before the first minute.