Microsoft watched an attacker's AI agent go from a key left in public code to stolen company data in about five minutes. OpenAI raised $11bn, at a lower valuation than it first asked for, and Canva is restructuring 22% of its staff after its AI features cost more to run than it planned. Further down, a ChatGPT sentence hidden in France's top prize-winning novel, and a ring that approves your AI agents.
~5 minutes
from a leaked GitHub secret to stolen Key Vault data, in one attack Microsoft tracked
$11bn
OpenAI's first close, at a $1.25 trillion valuation
1,271
Canva staff affected by its restructure, about 22% of the company
IN THIS ISSUE · 5 SECTIONS
1 Major AI News: AI agents abuse leaked keys, OpenAI's $11bn round, Canva's restructure.
2 Fun AI Topics: a ChatGPT sentence in France's top novel, video calls with Meta's AI, a ring that approves your agents.
3 AI Tools: ChatGPT Images 2.6, Gemini Omni 1.2 Flash, Inkling 2 Small, beacon.sh, Wanaka Studio.
4 Expert Prompt: find where your app's keys could leak.
5 Trending: AI spending goes flat, Perplexity Teammates, GPT-6 Cyber, DeepMind's RSA find, Heidi Health's cuts.
Forwarded this email? Subscribe here →
THE BIG STORIES
1 Major AI News
LEAD STORY
An attacker's AI agent went from a leaked key to stolen data in five minutes
Microsoft Threat Intelligence says attackers are using an AI agent framework called ARTEX to search public GitHub repositories for passwords, app credentials and AI service keys left in code. Microsoft has seen large volumes of these searches since August, and from 28 October the agents began signing in with what they found, across dozens of apps in several companies' Microsoft accounts. In one case an agent went from searching GitHub to pulling data out of an Azure Key Vault in about five minutes. When a request failed, it corrected its own mistake and got through on the next try.
Why it matters: Microsoft's five-minute example shows how little time a team now has between pushing a key to a public repo and someone using it. Nobody on the team needs to have looked at the commit yet.
Turn on secret scanning and push protection for every GitHub repo you own
Treat any key that was ever public as stolen, and rotate it even if the file has been deleted
Move app logins to managed identities where your cloud supports them, so there's no stored password to leak
ALSO THIS WEEK
OpenAI raises $11bn, at a lower price than it wanted
OpenAI has closed the first $11bn of a new funding round at a $1.25 trillion valuation, Bloomberg reported on 6 November, with SoftBank and MGX among the investors. It had pitched a higher figure, reported at between $1.4 trillion and $1.5 trillion, and early talk put the round at up to $30bn. A source told Bloomberg it could now reach about $20bn. TechCrunch reports Anthropic's expected IPO valuation has slipped too, from about $2 trillion in early October to $1.2 trillion to $1.3 trillion.
Why it matters: Both of the biggest AI labs are raising at lower prices than they wanted a month ago. Investors are still writing very large cheques, but they have started to negotiate.
The round is still open, so the final total could change
OpenAI also cut its fundraising goal by about $10bn, according to TechCrunch
ALSO THIS WEEK
Canva restructures 22% of its staff as AI costs bite
Canva announced a restructure on 4 November that affects 1,271 people, about 22% of its workforce. Not all of them will lose their jobs. Some will change roles or move to central hubs, and Canva hasn't said how many roles will go. In August it cut its 2026 revenue growth forecast from 30% to 20%, after the cost of running its AI features came in higher than expected. Chief executive Melanie Perkins called it the hardest moment in Canva's 13-year history.
Why it matters: Adding AI to a product people already pay for costs real money to run, and a company Canva's size is now adjusting its whole business around that bill.
Canva has not directly linked the restructure to its AI costs
Perkins is holding a company-wide meeting on 10 November to explain the changes
FOR THE COFFEE BREAK
2 Fun AI Topics
France's top book prize went to a novel with a ChatGPT sentence inside
Philippe Jaenada won the Prix Goncourt on 3 November for L'Inconnue du quai de Javel, then told France Inter he had used ChatGPT three times while writing it. One was a sentence written by ChatGPT that he slipped into the book to "play with the reader", and he only told the jury after the vote. The other two uses were for technical details. The Académie Goncourt's president, Philippe Claudel, says the rules will change before the end of the year to state that the prize goes to a work written by a human.
Why it is interesting: France's most famous literary prize went to a book with a machine-written line in it, and the jury only found out after it had voted.
Another novel was dropped from this year's longlist over suspected AI use
Jaenada says he also used ChatGPT's odd technical answers as material to poke fun at
Source: Euronews
You can now video-call Meta's AI, and its face reacts to you
Since 5 November, anyone in the US can phone or video-call Muse, Meta's AI assistant, on mobile, web and Mac. On a video call a cartoon avatar changes its expressions and gestures to match the mood of the conversation, and reacts while you talk so you can tell it's following. It can't see you. It only hears what you say. You can choose voice only, a small avatar beside the chat, or full screen.
Why it is interesting: Talking to an AI starts to look like a FaceTime call, which is a very different habit from typing into a chat box.
Voice is coming to Meta's AI glasses and the Muse Charm wearable, with no dates yet
It runs on two new models, Muse Realtime Voice and Muse Realtime Avatar
Source: YourStory
A $399 smart ring that approves your AI agents with a flick
Aina opened preorders on 5 November for a smart ring with a 0.9-inch screen, a speaker, a microphone and a small mechanical controller. It learns your routines from your phone, calendar and Mac, then offers actions when they're useful, such as booking a ride as you leave for work. You approve or reject requests from AI agents with a flick of the finger, mute meetings on your Mac, or hold the button and talk to it. Early-bird pricing starts at $399, and shipping is planned for early 2027.
Why it is interesting: As agents ask permission for more and more things, one company thinks the answer is a button you wear.
The battery lasts up to 18 hours
It pairs with iPhone and works with Mac
Source: Aina
TRY THIS WEEK
3 AI Tools
[IMAGES] ChatGPT Images 2.6 - Creates, edits and animates images in ChatGPT, now with GIF output and faster generation. openai.com (launch announcement)
[VIDEO] Gemini Omni 1.2 Flash - Makes and edits video clips up to 15 seconds long, at $0.07 a second for 720p, 30% cheaper than the last version. blog.google (launch announcement)
[MODELS] Inkling 2 Small - Thinking Machines' new model, ready to try and fine-tune on Tinker now, with open weights promised under Apache 2.0. thinkingmachines.ai (launch announcement)
[CODING] beacon.sh - Records your AI coding sessions across Claude Code, Cursor, Codex and 20+ other tools, and turns past fixes into notes that future agents reuse. github.com (open-source project)
[GAMES] Wanaka Studio - Describe a 3D game, play a first version in your browser, then refine it with AI or by hand. wanaka.app (product page)
COPY AND PASTE
4 Expert Prompt
Context: Microsoft's report shows agents hunting GitHub for keys left in code. Before you can rotate anything, you need a list of what could be exposed. Paste the names of your services and the files their settings live in, never the keys themselves.
You are my security checklist assistant. I will paste a list of the services, APIs and cloud accounts my [app or business] connects to, and the files or tools where their settings are kept. For each one, tell me what kind of secret it uses, where that secret is most likely to leak (code, config files, build logs, shared docs), how to check whether it was ever pushed to a public repo, and how to rotate it. Do not ask me to paste any real keys or passwords. If a service offers a login method that needs no stored secret, flag switching to it as the first step. Finish with the five riskiest items, ranked.
Example use case: A two-person software team lists Stripe, OpenAI, AWS and Supabase, learns to search its commit history for an old AWS key, and rotates anything it finds the same afternoon.
WHAT PEOPLE ARE TALKING ABOUT
5 Trending AI Topics
Company AI spending went flat over the summer
Ramp's spending data shows business spending on AI tokens grew just 1.5% across August and September, after a 42% jump from June to July. It picked up again with an 8.9% rise from 23 September to 23 October. The median company spends $12 to $15 per employee a month on AI, while the top 10% spend $600 to $1,000. Ramp's economist Ara Kharazian puts the slowdown down to price cuts, not weaker demand, since companies keep using more tokens while paying less for each one.
Why it is important: AI is getting cheaper for the businesses buying it. That makes the labs' trillion-dollar valuations harder to justify.
Source: TechCrunch
Perplexity's Teammates work in the background around the clock
Perplexity launched Teammates on 6 November, personal AI agents that work on ongoing tasks 24/7 and alert you when a decision is needed. You give each one a role, run several at once, and they can manage or add other Teammates as the work grows. Each has its own sidebar entry where you can check progress or redirect it. It needs a Perplexity Max plan at $200 a month, on web and iOS.
Why it is important: Perplexity is selling agents you hand whole projects to, and at $200 a month it's aimed at heavy users for now.
Source: Engadget
OpenAI gives GPT-6 Cyber its highest risk rating yet
OpenAI released GPT-6 Cyber on 5 November, a version of GPT-6 Astra trained for authorised security work such as finding software flaws and testing websites for weaknesses. Only approved members of its Daybreak Red programme can use it, through the API and Codex. OpenAI says it has already helped find thousands of vulnerabilities in widely used software, including Linux and Firefox, and rates it "Critical" for cybersecurity under its own safety framework, up from "High" for the previous version.
Why it is important: In the same week Microsoft showed attackers using agents, OpenAI shipped one for defenders and kept it behind a vetting process.
Source: Tech Insider
Google DeepMind's agents found two weaknesses in old 1024-bit encryption
Google DeepMind said on 6 November that AI agents, guided by human cryptography experts, found two new weaknesses in 1024-bit RSA, an older encryption standard still used in some legacy systems. One reportedly cuts the average cost of cracking each extra 1024-bit key by about 30 times when many keys are attacked together. DeepMind says modern 2048-bit RSA is not affected by either finding.
Why it is important: Old encryption is still running in plenty of older systems, and AI just made attacking it cheaper.
Source: Google DeepMind on X
AI scribe unicorn Heidi cuts 18% of staff two months after its big raise
Australian AI medical scribe maker Heidi Health is cutting about 18% of its workforce, at least 108 of its 600-plus staff, with the cuts falling mostly on its Melbourne R&D, support and sales teams. Two of its three co-founders, chief technology officer Yu Liu and chief financial officer Waleed Mussa, are also leaving. Heidi became a unicorn two months ago with a US$100m Series C. Chief executive Tom Kelly told staff the company had "made difficult choices about where to focus our team and investment".
Why it is important: Fresh unicorn funding didn't stop a restructure, and Heidi and Canva, two of Australia's best-known tech companies, both cut staff in the same week.
Source: Startup Daily
That's it for this week's Round-up. Forward this to one person who wants to stay ahead of AI. See you next week. - Nicolas
