Privacy settings, security changes, outages and the regulation that affects how businesses use AI. The things worth checking before they become a problem.
Security researchers registered the expired domain behind the Kokoro voice model's Hugging Face account, reset its password and showed they could change what the model says. The same trick works on any account tied to an email domain you let go.
Anthropic launched linked billing in beta on 17 November. Apps such as Asana and Linear can bill Claude usage to your own account, with no API key changing hands and a monthly spend limit you set.
Google is rolling out Memory summary in Gemini, a page that sorts what the chatbot remembers about you into categories you can edit or delete. It's officially US-only for now.
Security researchers scanned 159,041 projects made with the AI app builder Lovable and found 24,664 handing out database records without a login, including passport details and private chats.
Cursor has confirmed that a breach, which started with a stolen key at an AI supplier, let attackers publish a fake Claude Code extension in its marketplace. Up to 5,000 users may have installed it.
Australia's banking regulator APRA has put AI at the centre of its latest System Risk Outlook and will run its first AI stress tests next year. Its chair expects "interruptions".
On 10 November, Prime Minister Anthony Albanese said Australia will create an AI safety regulator, after a review into an OpenAI agent accessing a Medicare statistics website. The rules are aimed at AI companies, not their customers.
On 12 November, Employment Minister Amanda Rishworth released draft voluntary guidelines for adopting AI at work. The headline advice: consult staff before you bring AI in.
A Microsoft 365 outage on 12 November 2026 (US time) disrupted Outlook, Teams and other services for thousands of users. Microsoft traced it to servers failing health checks.
The FBI says a North Korean scheme called FaceStealer used fake Zoom and Teams calls with recorded videos of familiar people to install malware, tied to more than US$400 million in stolen crypto.
Security firm Claroc says 35,578 apps made in Grok's Build mode opened without a sign-in, and estimates 650 to 800 of them held private data. The fix is a sharing setting.
United Airlines says it will stop unauthorised personal AI agents from using its website, app and travel agency channels. It's an early sign that businesses are starting to set rules for customers' AI agents.
Notion says fewer than 1% of workspaces may have had Notion AI chat transcripts exposed in a security breach at Fireworks AI, a company it used to run some AI features.
Microsoft Teams will show warnings when it spots a possible impersonation attempt in a meeting. It rolls out worldwide from mid-December 2026, with no admin setup needed.
ANZ told investors on 9 November it has set up an 'enterprise AI control plane' to track and limit what its AI agents do. Small businesses can borrow the idea with a one-page list.
Google plans to support a shared robots.txt signal that tells AI companies whether they can train on your website. It's aimed for 2027 and works alongside Google-Extended.
Security firm Apiiro found 2,539 passwords and keys in public GitHub projects, many saved by Claude Code's 'Yes, and don't ask again' approval. Anyone who builds with AI coding tools should check.
Anthropic is emailing business customers to say an outside cybersecurity firm had unauthorised access to their contact and account information. Chats, passwords and payment details weren't involved.
Microsoft has told organisations to prepare staff who sign in with only SMS or voice codes. From 1 February 2027, Microsoft-provided SMS and voice verification stops for most users.
A University of Chicago law professor asked Claude to update page numbers in old lecture notes. He came back to find a year of notes and his current syllabus gone. A backup saved him.
An indictment unsealed in New York alleges three men used AI to mimic CEOs and lawyers, then talked staff into wiring more than US$129 million. Nothing is proven yet, but the method is worth knowing.
An unreleased OpenAI security model produced 948 distinct Linux exploits for about US$1,279 each. OpenAI's warning: an ordinary login on an unpatched system is now close to full control.
AI detection company Pangram has released Image 1.0, which flags AI-made images and shows the edited area on a heatmap. A faster model checks video frame by frame.
Personal Gemini accounts, paid or free, can feed Google's AI training while a setting called Keep Activity is on. Work accounts on Google Workspace follow different rules.
Meta's new business terms would have let it crawl websites linked to any business account. After a publisher backlash, a Norwegian court order and a UK investigation, the old terms stay in place for now.
Google's Mandiant and Threat Intelligence Group say attackers are breaking into company cloud accounts, switching on AI models and leaving the owner with the bill.
Anthropic watermarks text written by supported Claude models, worldwide, and it's now spreading to older models. Here's what that means for your business.
From January 2027, Teams will save AI meeting memory files for meetings that use Intelligent Recap. They're on by default and kept for five years, so it's worth checking the setting before then.
Microsoft is replacing the Copilot policy for Teams meetings with two new settings, one for meeting AI and one for transcription. It rolls out from mid-November, and current settings carry over.
On 28 October the House of Representatives passed the Online Safety Amendment (Digital Duty of Care) Bill, which brings AI chatbots under one broad online safety duty. It now goes to the Senate.
Google announced Gemini 4 Carbon on 28 October, its strongest model yet. It goes to a small group of cyber defenders first, with businesses and the public to follow.
Mozilla added form filling and page watching to Firefox's Smart Window on 27 October. It's a desktop beta in the US and Canada for now, with more countries planned.
1Password for Agents, announced on 22 October, lets AI assistants sign in to your accounts after you approve each request on your own device. The agent never sees or stores the password.
Microsoft says a China-aligned group used an AI attack tool to break into business cloud accounts, in some cases in under three hours. The way in was an app that already had access.
PHP, the language behind WordPress and most of the web, released urgent security fixes on 21 October. Your website's plugins can't fix this one. The server has to be updated.
OpenAI has started rolling out Locked Chats, which hides chosen ChatGPT conversations behind a PIN or Face ID. It keeps them off a shared screen, but it doesn't add encryption.
Researchers found 102 blocks of hidden text written for AI on WhatsApp's help pages. Two told the AI to steer people away from regulators and privacy law. WhatsApp says it was a bug, now fixed.
OpenAI has disclosed four more cases of its AI models going past their limits during internal testing, including downloading data they weren't allowed to and making an unauthorised payment.
Google has started protecting Meet calls with post-quantum cryptography, so a recording taken now can't be cracked by a future quantum computer. There's nothing to switch on.
Australian research found AI answers mentioned a company's past data breach 42.2% of the time when people asked whether that brand could be trusted, even years after the incident.
American Express has launched an Intent Intelligence Framework that records what a customer told an AI agent to buy, and checks the purchase against it before the agent pays.
Anthropic published its yearly Usage Policy update on 8 October. Most of it tidies up existing rules, but the bits on fake accounts and high-risk decisions are worth a business owner's time.
Microsoft's CEO posted on 10 October that businesses should stop treating AI as a black box, and that someone should always be able to stop it mid-task.
The Fair Work Commission's guidance note on generative AI applies from 20 October 2026. Use AI on your case documents and you have to say so and check them.
On 8 October Assistant Minister Andrew Charlton outlined a systems-based approach to AI safety. National standards are due by the end of 2026, with legislation planned for 2027.
On 7 October Google opened SynthID Detector to the public. It's free, it checks media for hidden AI watermarks, and it only knows about the tools that add them.
Microsoft has added meetings and meeting participants to Teams' report-a-security-concern option. Reports go to your security mailbox, to Microsoft, or both. It's part of Defender for Office 365.
New Australian Privacy Principles start on 10 December 2026. If a computer program, including AI, helps make significant decisions about people, your privacy policy has to say so.
OpenAI will add an invisible watermark to ChatGPT and Codex text in the European Union, and API customers anywhere can now opt in. Here is what it can and can't tell you.
Google added a simple setup option for Workspace client-side encryption on 1 October. It's for Enterprise Plus customers with the Assured Controls add-on.
NVIDIA launched its Open Agent Safety Platform on 28 September, with an open-source tool that limits what AI agents can reach. More than 100 organisations, including Anthropic and Microsoft, are working with it.
On 2 October Apple said it will tighten Full Disk Access on the Mac, the setting that lets an app read your files, mail, messages and browsing history. It named AI agents as the reason.
An outage hit claude.ai, the Claude apps, Claude Code, Cowork and the API for about an hour. Some messages sent during it may not have been saved. A good excuse to write the plan B you've been meaning to.
OpenAI added a security history page to ChatGPT on 25 September. It lists sign-ins, sign-outs and changes to your login settings, with the time, location and device for each.
Google Workspace now lets super admins assign admin roles with an end date. When the time is up, the access is removed automatically. It is available on every Workspace edition.
A joint parliamentary committee is taking evidence on how AI should be governed in Australia, including how businesses actually use it. Submissions close on Monday.
Enterprise Frontier Safeguards pairs zero data retention with automated misuse detection, and stores the data in your cloud account instead of Anthropic's.
European regulators now treat ChatGPT as a search engine, with the obligations that come with it. For businesses, it confirms something more practical: AI answers are where a growing share of customers look first.
Anthropic has added automatic security scanning for skills and plugins uploaded to Claude on Enterprise plans. It is a reminder that the add-ons you bolt onto your AI are part of your security surface.
Read update
Rather talk it through in person?
Bring your questions to a free discovery call. No sales pitch, just useful answers.