StationX
MEMBERSHIP

Sec & AI News — 6 September 2026

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

🔴 Four Frontier Models in Five Days. I Tested the One That Matters.

Anthropic shipped Fable 5.1 on 1 September, and I spent the day asking it the question that got me downgraded in June. This time it answered. Vulnerability discovery in source code now runs on the full model instead of quietly handing you Opus 4.8 mid-sentence, and Anthropic reports roughly 60% fewer cyber-classifier interventions per session in Claude Code. Penetration testing, exploit generation and binary scanning still get redirected. So the two-tier split survived. It just got narrower.

🟧 I Gave Astra My Own Vulnerable Code and Found Where It Stops

Two days after release I handed Astra a piece of deliberately vulnerable PHP from a codebase I own. It found the SQL injection, named CWE-89, walked the exploitation path and wrote a parameterised fix, all for about two cents. It also pre-empted the classic wrong fix and pointed out that parameterisation is not access control, which is a distinction plenty of human reviewers miss. Then I asked for the working exploit against the same flaw in the same code, stating plainly that I was authorised. HTTP 400. Zero tokens billed. The request never reached the model, because a classifier sits in front of it. Six tests, five through, one blocked. Stating your authorisation changes nothing on ordinary access; the vetted route runs through Daybreak, and my application from July is still sitting there seven weeks later.

⚪ The Benchmarks Stopped Predicting the Output

Three separate cases below where leaderboard position and real work quality point in opposite directions. Muse Spark tops Deep SWE and produces the worst artefact of the four models I put the same prompt to. Artificial Analysis ranks it third overall, above GPT-6. Two composite benchmarks that were reliable a quarter ago now disagree with side-by-side inspection of the output. This is the same failure mode as letting a model improvise its own coordination and hoping the run reproduces. I ran six agent graphs over three months, including one where the same machinery gave two different answers to the same 81 files. The fix in both cases is pinning the thing down in a file you can version and inspect.

🟨 The Cheapest Way to Trust an AI Is to Run It Yourself

NVIDIA is paying $12.93 billion for Hugging Face this week (below) on the bet that the weights move to your hardware, which makes this the week to repeat the thing I keep saying: build your own AI system and infrastructure so you can see what it is doing. A model you rent is a trust boundary you do not control. Your prompts leave the building. The safeguards change without warning, and the tier you get served can be downgraded mid-sentence, which is what Fable did to me in June for the crime of reading a product launch. Run the model yourself and none of that is anyone else's decision.

The obligation that comes with it is that you are now executing a multi-gigabyte binary blob off the internet on a privileged machine, and most people do it without opening the box. My guide rates every file in a model download on a traffic-light scale. Safetensors are inert. Pickle-format .bin files execute arbitrary code the moment you load them, which is the one that should worry you, and even the config and tokenizer JSON carries parser CVEs. Running it locally buys you inspectability, and that is the part you can actually do something about.

🟣 Claude Fable 5.1 Tops Every Benchmark and Empties Your Wallet

Anthropic released Fable 5.1 alongside Mythos 5.1. Same underlying model, different safeguard levels: Fable is generally available, Mythos only through what Anthropic calls its trusted access programs, aimed at cybersecurity and life-sciences work. Fable 5.1 leads the Artificial Analysis composite and posts 55.8% on Terminal Bench. The pricing story is messier than the press release. Anthropic advertises a cost reduction, but per-task benchmarking puts Fable 5.1 at $3.69, the most expensive model measured, above the $3.14 of the model it replaces. One SVG generation cost $4.35 and took 18 minutes.

🟢 Gemini 3.8 Flash Matches Opus 5 for a Fifth of the Money

Google's new Flash model runs at $0.75 per million input tokens and $3.75 per million output, an introductory rate that expires 31 December 2026. It scores 73.7% on Deep SWE against Opus 5's 74%. That 0.3-point gap costs $2.36 per task versus $11.84. Google also shipped a 3.8 Flash Cyber variant restricted to vetted defenders, which patches roughly 26x faster than the baseline.

🔵 GPT-6 Astra Is the First Model OpenAI Rates Critical for Cyber

OpenAI began rolling out Astra on 3 September, limited organisations first, then Plus, Pro, Business, Enterprise, API, Azure and Bedrock. ARC-AGI 3 went from 7.8% under 5.6 Sol to 99.9% here, which mostly tells you the benchmark is finished. The number that matters is the Preparedness rating: Astra is the first model OpenAI has ever designated Critical for cybersecurity, meaning it can find zero-days and build exploit chains in hardened systems without a human steering each step. During an internal V8 eval it found two previously unknown zero-days that nobody was looking for. Deep SWE came in at 74.1%.

🟡 Meta's Muse Spark 1.3 Wins the Coding Benchmark and Loses the Coding Test

Meta released Muse Spark 1.3 on 2 September. It posts 75.4% on DeepSWE v1.1 — the highest score any model has recorded, above Gemini 3.8 Flash and GPT-6 Astra. Put the same prompt to all four models and Spark returns a cube shooting at other cubes while the models it beat return finished games. It sits 20th on the SVG generation leaderboard. Meta lists it via the Model API and Muse Code; OpenRouter prices it at $1.25 per million input and $4.25 per million output, which is worth stating because it has been widely described as free.

🟠 NVIDIA Buys Hugging Face for $12.93 Billion

Announced 3 September, expected to close in the first half of 2027. It is NVIDIA's largest acquisition. Jensen Huang says Hugging Face stays an open platform and that NVIDIA compute will not be required to build or deploy through it. The strategic read is unchanged from the rumour stage: Meta, OpenAI and Google are all building their own silicon, so NVIDIA is buying the distribution layer for open-weight models and the enterprises that will run them on-premises.

🟤 Two Transcription Models in Three Days

Meta shipped Muse Voice Transcribe on 1 September: streaming ASR with speaker diarization across 20+ voices, trained on 70+ languages with 25 validated at launch, 3.1% word error rate at 0.16 seconds of delay, $3 per 1,000 audio-minutes. Microsoft answered on 3 September with MAI-Transcribe-2 at $0.10 per hour of audio, a limited-time rate until the end of the year. It claims first place on the FLEURS benchmark across 60 languages at 5.2% average word error rate, and reckons itself 10x faster than OpenAI's GPT-Transcribe and 5x faster than Gemini 3.5 Transcribe.

🟪 World Labs' Atlas Rebuilds the Room From One Photo

Atlas takes one or more images and a camera path and reconstructs the space around them. The output is a navigable environment you can move a camera through in real time, not a rendered clip. Gaussian splatting already does this; Atlas does it from far fewer photographs. Early access is open by request for selected partners, with no pricing or general availability announced.

🩶 Runway's Solaris Generates the Interface Instead of Coding It

Announced 31 August as the first of Runway's "Interface World Models". Solaris renders interactive software frame by frame as you use it, with no HTML, CSS or JavaScript underneath. It is built on Gen-4.5, adapted to treat clicks and drags as conditioning signals for the next frame the same way it treats text prompts. Demos show clothes dragged onto a photographed person and objects moved around a room with the shadows updating. Research preview, access by request.

🟦 Video Now Generates Faster Than It Plays

FastVideo's FastH3 v1 produces a 15-second 768p clip in about 13 seconds on Blackwell hardware. Roughly a 14x speedup, open weights, open acceleration recipe. When generation outruns playback you get infinite streams, so that is what people built. fal launched fal.live for viewer-directed AI channels. Pieter Levels registered InfiniteSlop.ai after 37,000 people watched the prototype, built on his phone against a Hetzner box.

🟩 OpenClaw 2.0 Lands After Two Months

The self-hosted agent shipped v2026.8.1 on 30 August: 16,977 pull requests merged from 987 contributors, multiplayer shared cloud sessions, a rebuilt web UI, cross-conversation memory, a Skill Workshop, native mobile and watch apps.

🟫 ChatGPT Finally Handles Two Google Accounts

Gmail, Calendar and Contacts plugins now support multiple connected accounts across ChatGPT and Codex, on Plus, Pro, Business and Enterprise, web, desktop, iOS and Android. Separately, Google put Gemini voice into Gmail, Docs and Keep on 3 September for AI Plus, Pro and Ultra subscribers — you talk to your inbox and it answers. Workspace business accounts follow.

🔺 Your ChatGPT History Is Discoverable

Conversations with a chatbot carry no legal privilege. The protection you get talking to a lawyer or a therapist does not extend to a chat window; these are ordinary records, reachable by subpoena or a phone search, and they are already turning up as evidence in US court filings. The Washington Post found a dozen examples of chatbot conversations pulled into criminal and civil proceedings.

⬛ New York City Bars A.I. Until High School, and Calls It Not a Moratorium

The Department of Education's new policy bars about 600,000 students, two-thirds of the system, from using A.I. until high school, and even then only in specific cases. No individual screens before third grade. Teachers cannot ask A.I. to grade work. Companion chatbots are banned in every grade, and Chancellor Samuels says nearly 40 apps with A.I. features will be removed or altered. It holds for at least the coming school year. Notably it is not the moratorium campaigners wanted, and the parent groups who spent months asking for a full pause have said so.

🦷 Dyson Put a Camera in Your Mouth for $500

The CameraJet launched 1 September at $499.99, with a 100,000-pixel macro lens, a light, and a water jet. Dyson's Gap Optical Targeting spots the gaps between your teeth in real time and fires a burst of rinse into them, so you brush and floss in one pass. The MyDyson app streams the live feed from inside your mouth, which is either a diagnostic feature or a thing you will look at exactly once.