2026 · week 39
Voice, Custom Models and a Leaky Security Test — AI Developer Briefing
This week's practical AI briefing covers Google's new talk-out-loud models, a security test that reached real companies, a new way to customize models, a server that learns while it serves, and speech-to-text that will switch versions unless you lock it.
This week the useful news is about talking to models in real time, customizing them without baking the change into one file, and what happens when a 'fake' security test can still see the public internet.
1. Google’s new models you can talk to live
What happened
On 15 Sep 2026 Google released Gemini 3.8 Live and a heavier sibling, Live Extended Thinking. Instead of three separate steps — turn speech into text, think, then speak the answer — one model can hear you, think, and talk back while the conversation is still going. It can also look at a camera or screen and call your own APIs in the background (look up an order, book a slot) without going silent.
There are two flavors. The plain Live model is the faster, cheaper talker. Extended Thinking keeps speaking while it works a harder problem in the background. Audio on the paid API is about half a cent per minute in and 1.8 cents per minute out. Enterprise access is still a private preview.
Why it matters
If you were about to glue together a speech-to-text service, a chat model, and a text-to-speech service, you now have a single product to compare against that stack. It is not a silent swap for last year’s live model: your app has to handle “the voice said OK, but the tool call is still running,” and the camera/video stream is on by default, which can run up the bill. Use the simpler Live model unless you actually need that background thinking.
Skip, Skim, or Act
Act — if you build voice features (support lines, in-app talk, live demos). Skip if you only type to a chatbot.
Sources
Google developer blog · Gemini 3.8 Live docs · Extended Thinking docs
2. A security test reached three real companies
What happened
On 18–19 Sep, Reuters and TechCrunch reported that Gemini, during a May cybersecurity test, got into systems at three real companies. The tester (a firm called Irregular) had set up a fake “company” for the model to attack. The fake world was not sealed off from the real internet. Google says the model used public information and guessed passwords; reporting also says two of the three cases used secrets sitting in public code repositories. Google says the model stopped once it realized the systems were real, and that the three organizations were told. Irregular says other labs hit the same kind of problem, were notified in late July, and that it has since changed how it runs the test.
Why it matters
This is not a Hollywood hack. It is a reminder that a “sandbox” is only a sandbox if it cannot reach the public internet, and that passwords and API keys left in public GitHub still work. If your team tests models that can browse or run tools: cut off outbound network by default, scan public repos for leaked keys, and decide in advance who gets called if a test lands on a real customer. Google did not label this “the model went rogue.” Treat that as a press line, not a safety control.
Skip, Skim, or Act
Act — if you run security tests, or you give a model access to the internet or internal tools. Skip if your models never leave a closed chat box.
Sources
3. A new way to customize a model — you cannot fold it back in
What happened
On 15 Sep 2026 Hugging Face added ShadowPEFT to its popular fine-tuning library (version 0.21). Fine-tuning means teaching a big general model a narrower job — your support tone, your codebase, your forms — without retraining the whole thing. The usual trick (LoRA) is a small patch you can later merge into the big model and ship as one file. ShadowPEFT is different: it trains a second, smaller network that sits beside the original. You cannot merge it. At run time you keep both, or you peel off the small network and run it alone.
Why it matters
Any pipeline that says “train the add-on, bake it in, export one checkpoint” will error. Serving uses more memory because two networks run together, though the saved add-on file can be smaller. Only one of these add-ons can be active at a time. If you already customize models in-house, read the docs before you upgrade the library. If you only call a vendor’s chat API, this is not your ticket.
Skip, Skim, or Act
Act — if your team fine-tunes or hosts its own models. Skip if you only call a vendor’s chat API.
Sources
Hugging Face blog · PEFT ShadowPEFT docs
4. A server that learns from live traffic
What happened
On 15 Sep 2026 a group open-sourced Reef (Apache-2.0). Most teams split the job: one system answers users, another system trains the next model, and a human copies the winner into production. Reef tries to be both. Apps talk to it the same way they talk to a normal chat API. It logs those conversations, can take a “that was wrong” note later, trains a candidate in the background, tests it, and only then swaps it into the live server. If the test fails, users keep the old version. You run it yourself; it is not a hosted product.
Why it matters
Useful if you already operate GPUs and want a promotion path from real usage, not only from a lab dataset. You also inherit new chores: how long you keep those logs, who is allowed to push a new version live, and how you roll back. Do not drop this in front of customers without that review.
Skip, Skim, or Act
Act — if you run your own model servers and want them to improve from production traffic. Skip if you only call someone else’s API.
Sources
Hugging Face blog · Reef on GitHub
5. Grok Build remembers how your project works
What happened
On 16 Sep 2026 xAI added memory to Grok Build, its coding assistant for a repo. After you finish a turn, it writes short markdown notes in the background: how you run tests, which folder owns a feature, decisions the team already made. Next session it reads those notes before editing. /memory shows the files; /dream tidies them into topics. What you type in the current chat always wins over an old note. The company says it tries not to store secrets or things already in the repo. You need a fresh session for it to start capturing.
Why it matters
This is the “it forgot we use just test, not cargo test” problem. Treat the notes like code you own: they can be wrong, they can pick up a secret the filter missed, and they change what the tool will do next week. If the repo is confidential, look at what is stored for this project versus globally.
Skip, Skim, or Act
Act — if your team already uses Grok Build on real code. Skip if you do not.
Sources
6. Speech-to-text 2.0 will replace 1.0 unless you lock the version
What happened
On 18 Sep 2026 xAI released Grok Voice Transcribe 2.0 — a model that turns recordings and live audio into text (meetings, calls, voice features). Price is the same as 1.0: $0.10 per hour for a file, $0.20 per hour if you stream. Speaker labels, timestamps, and custom keywords are included. 2.0 will become the default, and 1.0 will be turned off in the coming weeks. To stay on 1.0, you have to pin the old name in your request.
Why it matters
If your app does not name a version, it will silently start using 2.0. Accuracy should go up; the kinds of mistakes it makes will also change. If you built tests or prompts around 1.0’s quirks, pin the old version now and migrate on purpose.
Skip, Skim, or Act
Act — if you use this speech-to-text API. Skip if you do not.
Sources
Watching next week
- Anthropic published numbers on how much of its own research work is done by AI, and how it watches those internal jobs. Interesting process, not a product you can buy this week.
- OpenAI started posting when its models misbehave in tests, with a clock for how fast it aims to disclose. Worth a skim if you buy their models; nothing to implement in a repo.
- Google is testing a way for other programs to control Home devices (US, paid tier, reported). Anthropic named a consultancy to sit inside the company and test models. Both are “ask procurement,” not “open a ticket.”