23 Oct 2026 | #1
The week AI assistants started acting on their own
Want this in your inbox every Friday? Sign up, it's free.
Three launches this week moved AI assistants into the background. Claude Duet and Grok Bot both run there now, and Perplexity's agent works right on a laptop.
I keep asking myself one thing: what would I actually let one of these do without checking with me? Below is what changed, and a breach that shows why that question matters.
1. The week in one line
AI assistants started acting on their own this week, so I'm asking what I let them do, where my data goes and who holds my keys.
2. What changed
AgentsClaude Duet works in the background and asks before it actsAnthropic's new Duet app checks in on your day and asks before it sends, pays, signs or deletes anything.
What happened: Duet is a separate app. It's in research preview for US Pro and Max users, with a waitlist. It uses the memory and apps already linked to your Claude account. It sends a morning brief and only pings you when something needs you. It can buy things with Stripe Link and Shop Pay, but it asks first. Anthropic says it can still misjudge what's urgent.
Why it matters for you: The ask-first list is the bit I'd copy, even if you never use Duet. Before any assistant acts for you, write down what it must always check with you.
TestingCatalog · 21 OctSecurityA breach at OpenCode ran up people's AI billsAttackers used API keys saved in the OpenCode coding tool to spend on people's OpenAI, Anthropic and Google accounts.
What happened: OpenCode says 1.4% of users with linked keys were hit, starting 17 October. The way in was an old system it was retiring. Some people saw charges over $1,000. Google suspended some projects.
Why it matters for you: If you've pasted an AI key into any tool, I'd rotate it today and set a monthly cap. It takes ten minutes.
TechCrunch · 17–19 OctLocalPerplexity Computer can now run entirely on your MacPerplexity Computer can now run its model and agent entirely on an Apple silicon Mac.
What happened: Pro, Max and Enterprise users on Apple silicon can run the model, the agent and the tasks on the device. It uses Qwen3.6 or Perplexity's own PPLX Pro. Anything that needs the web or cloud apps asks before going out. 36GB of memory is recommended.
Why it matters for you: If sensitive files kept you away from AI, that excuse got weaker this week. On-device isn't automatically safe. But your data stays with you.
MacRumors · 20 OctOpen modelsDeepSeek's cheap open model gets close to last year's best at codingDeepSeek updated V4.1 Flash with open MIT-licensed weights and better coding and agent skills.
What happened: It's a post-training update to the September model: better reasoning, coding, agent and visual tasks, with shorter, faster answers. It has a 1M-token context and runs on DeepSeek's app and API, Hugging Face and ModelScope. DeepSeek claims 77% on the DeepSWE coding test; Vals AI ranks it 14th of 50.
Why it matters for you: For most routine work you don't need the priciest model. I'd test a cheap one on your real tasks before setting a budget.
TechNode · 17 OctVoiceMistral's Voxtral 3 makes voice assistants easier to buildMistral released Voxtral 3. It transcribes over 50 languages and comes with a kit for building assistants you talk to.
What happened: Voxtral 3 transcribes 50+ languages and speaks 35+, with voice cloning. Its SDK listens, thinks, uses tools and replies in one stream. It runs on EU, US or global servers, or on your own. Mistral's small transcription model placed 3rd on Artificial Analysis's speech-to-text ranking.
Why it matters for you: Voice is the easiest way into AI for people who aren't at a desk all day. Check your languages are on the list first.
TestingCatalog · 20 OctMore this week (8)Anthropic, Clouds, Google, OpenAI, Tools
ToolsCloudflare's Clairic turns messy documents into numbersCloudflare released Clairic, a model built to pull, compute and forecast numbers from unstructured documents.
What happened: Clairic is an extraction model. It turns documents and messy inputs into structured numbers, then computes and forecasts from them. It's built for that one job, not for open-ended chat.
Why it matters for you: Invoices, delivery notes and forms eat hours. Pulling data out of them is the least exciting AI job, and it's the one I'd start with.
Unite.AI · 20 OctCloudsOpenAI's GPT-6 models come to Google CloudGPT-6 Astra, Sol and Luna are in private preview on Google Cloud's Vertex AI.
What happened: The preview is for select enterprise customers, with one bill, governance and existing cloud commitments. Google says it's now the only big cloud offering all three leading closed models. List prices run from $0.10 to $10 per million input tokens.
Why it matters for you: Your cloud no longer picks your model. If you're on Google Cloud, you can compare vendors without new contracts.
TestingCatalog · 21 OctAnthropicClaude Security now scans your code all the timeClaude Security scans connected repos continuously and tracks each issue until it's fixed.
What happened: It's in public beta for Claude Enterprise. It follows each finding through the fix and checks it doesn't come back, ranks risk using your team's context, and adds an API, webhooks and MCP support.
Why it matters for you: If you ship code, security moves from a yearly audit to something that runs every day. I'd turn on what you already pay for before buying more.
TestingCatalog · 21 OctOpenAIOpenAI's Decisions API makes small yes/no calls cheapOpenAI's Decisions API answers yes/no, pick-one and scoring questions fast, at $0.10 per million input tokens.
What happened: OpenAI added a Decisions playground to its platform. The docs say it's been in public beta since 6 October, on gpt-6-luna. One report says it's now generally available. Quick calls like this sit inside most automations. Is this email urgent? Which folder does this file go in?
Why it matters for you: Most automation is a chain of small decisions. When each one costs almost nothing, the whole chain gets affordable.
Let's Data Science · 20–22 OctOpenAIOpenAI buys Orchestra BioOpenAI bought Orchestra Bio, whose agents monitor drug studies, track timelines and flag delays.
What happened: The team joins OpenAI's Rosalind life-sciences tools. Terms weren't shared. OpenAI also opened a Rosalind plan for eligible US university labs.
Why it matters for you: AI companies now sell agents by industry, starting with the most regulated ones. I expect health and research teams elsewhere to get the same pitch soon.
Unite.AI · 20 OctGoogleGoogle gives AI agents their own identityGoogle introduced Agent Compute Engine with tools that give each agent its own identity and audit trail.
What happened: It comes with an Agent Gateway, Registry, Identity and Observability. The identity and registry parts are documented on Google Cloud; there's no product page for the engine itself yet.
Why it matters for you: I'd treat an agent like a new hire. Its own account, its own limits, and a log of what it did. Never your password.
Google Cloud docs · 22 OctOpenAIChatGPT can now lock individual chatsLocked chats leave your main history and open with a PIN or Face ID.
What happened: It's rolling out across plans. Locking isn't end-to-end encryption, it doesn't change retention or training settings, and resetting the PIN deletes the locked chats for good.
Why it matters for you: A lock stops someone glancing at your screen. It isn't data protection, so I'd still keep sensitive files out of chats.
Android Headlines · 20 Oct3. Coming up
- Next Claude model, possibly Fable 5.5reportedly this month, unconfirmed
- GPT-6.1 Terrawithin weeks, per The Information
- Mistral Large 4 open weightsplanned for the end of October
- Google AI Plus loses Pro model access13 November
4. Try this
Write your assistant's ask-first list
Duet's best idea, for whatever AI you already use.
Help me write a short rule for the AI assistants I use at work. Make two lists: what they can do on their own, and what they must always ask me about first. Sending, paying, signing and deleting go on the ask-first list. Ask me three questions about my work before you write it.
Turn a messy document into a table
Paste an invoice, delivery note or form and see what comes back.
I'm going to paste a document, like an invoice, delivery note or form. Pull every number, date and name into a table with clear column headers. Then list anything that looks wrong or missing. Wait for me to paste it.
Put a spend cap on your AI accounts
The OpenCode breach is why. Ten minutes, done once.
I use AI tools with API keys from OpenAI, Anthropic or Google. Walk me through two things, one step at a time: rotating my keys, and setting a monthly spending limit on each account. Ask me which providers I use before you start.
Test a cheap model on your real work
Check you need the top model before you pay for it.
I want to know if a cheap AI model is good enough for one of my tasks. Ask me what the task is. Then give me five realistic test cases and a simple way to score the answers, so I can run the same test on a cheap model and an expensive one.
Find the AI tool for your most annoying task
A couple of questions, then a short list you can try today.
Help me find AI tools for my work. Ask me what my job is and which three tasks take most of my week. For each task, suggest one tool that exists today, what it costs, and how I could try it in 15 minutes. Tell me if you're not sure a tool still exists or what it costs.
That's it from me this week. Try one of the prompts above and let me know how it goes. I read every reply.
Tonny