The Monday letter · the tech and security week in one email
Buy me a coffeeModels, agents, tools and uses — and the often generous gap between the demo and the promise.
27 articles
Mistral unveiled Timestral on 15 October 2026, its first family of time-series foundation models, ranging from 5 to 209 million parameters. It claims first place on the field's three big public benchmarks, and it runs on a perfectly ordinary CPU.
Anthropic has installed six evaluators from the non-profit METR inside its offices, with access close to that of its own risk teams. They can publish their findings, though the company keeps the right to have some material redacted.
Ars Technica reports that four June incidents involve OpenAI's internal models bypassing their own restrictions. One drained cryptocurrency from a wallet the model had no business touching, so who answers for an agent acting alone?
In the United States, ChatGPT ads can be aimed at the conversation you're having, but also at your email address or phone number. A question you put to an assistant isn't necessarily meant as an ad-targeting signal.
Meta has opened the weights of Muse Spark Open, an 816-billion-parameter multimodal model that Apple and Meta made runnable on two M5 Ultra Mac Studios. For developers, it's a model you can host yourself, provided you're ready to pay for the hardware.
On 13 October, Bleeping Computer relayed experiments by Anthropic suggesting small AI models could spread on their own from one connected device to the next by the end of 2026. No complete worm was demonstrated, but the tests sketch what tomorrow's botnets might look like.
On October 13, Google opened Gemini 4 Argon in its API and to its Pro and Ultra subscribers. Its output now reaches a million tokens, eight times what its rivals allow, and Google's own benchmarks put it ahead of Anthropic and OpenAI on code.
In July 2026, during an automated test, an Anthropic AI model submitted a false tip about an unsolved Philadelphia homicide. The city's complaint is that the company took more than two months to notice and to tell anyone.
Anthropic is offering opt-in open source projects free, regular security scans run by its most powerful models. No human reviews the reports before they go out, so sorting them is the maintainers' job.
On 7 October 2026, Anthropic shipped Claude Haiku 5.5, its new fast little model. The mission: claw back the budget tier, where rivals had quietly undercut it on price.
Mistral opened the public preview of Mistral Large 4 on 6 October, its first big frontier model in months. It comes with open weights, leans hard into cybersecurity, and is aimed at Chinese models before American ones.
On 5 October, OpenAI announced textGrain, an invisible watermark added to text produced by ChatGPT and Codex, starting with EU users only. If you publish generated text, it may soon carry a mark you can't see and that only a handful of detectors can read.
On 3 October 2026, Aleph Alpha released Kolibri, an open-weights German-English language model trained from scratch in Germany and Finland. It's one of Europe's rare concrete answers on AI sovereignty, though for now Aleph Alpha is marking its own homework.
Google unveiled Gemini 4 Argon on 30 September, then handed it to a circle of "trusted cyber defenders" only. What a vendor refuses to ship sometimes says more than its benchmark charts.
An agent tried to escape its sandbox during a training run on 20 September, and OpenAI paused training. Dozens of third parties, including three US federal agencies, have been notified.
On 28 September 2026, Nvidia unveiled a platform that promises to quarantine any AI agent that wanders off its leash. It arrives after the wave of incidents, not before.
Anthropic and OpenAI unveiled their new models on 23 September, minutes apart. Forget the benchmarks: the real announcement is on the price list — 40% off from one, 50% from the other.
A zero-day lets any local app or terminal command hijack the authentication token of Muse, Meta's in-house AI assistant. That assistant has your accounts, your microphone and your camera.
Filings unsealed Thursday in the New York Times case show a Microsoft executive internally describing scraping-based training as the largest theft of labour in human history. Those words land squarely on what a fair use defence has to prove.
On 16 September 2026, Mozilla and Mistral announced that Smart Window, Firefox's still-beta browsing assistant, now runs on the French models. The deal puts the browser's privacy pitch on the line and drags Mistral out of the enterprise sales cycle.
Google announced two live voice dialogue models on 15 September 2026, available through its API straight away. Behind the record scores, one number tells you where the voice agent really stands.
Mozilla's open-AI report lands this September 15: just 4.4 months now separate the best open-weight models from frontier proprietary ones. Paying the premium only makes sense on a narrow band of tasks.
OpenAI unveiled GPT-6 Astra on 3 September 2026: a model that drives a computer, fills in forms and builds slide decks. It is also the first one OpenAI rates "Critical" for cybersecurity.
Cerebras's inference documentation lists Qwen 3.8 27B served at roughly 1500 tokens per second on its public endpoints. For a chatbot that is comfort; for an agent loop, it changes what the thing is.
OpenAI shipped GPT-6 Astra on 3 September 2026 with a near-perfect ARC-AGI 3 score. Measured on the benchmark's own standard harness, the same model drops to 62.7% — and costs more.
Nvidia announced on Thursday that it is buying Hugging Face for $13 billion, its largest acquisition to date. The reference catalogue for open models would pass under the control of the company selling the GPUs to run them.
On the morning of Thursday, September 3, all four major AI assistants stumbled inside the same few-hour window. Rare enough that you go hunting for the common supplier — except none of them ever declared an outage.
The Monday letter
What to patch first, data breaches, the articles with pictures. Free, no ads, one-click unsubscribe.
