Saturday 14 November 2026  •  By Timothy, CPA — Managing Director, Professional Financelink (PFL)

AI News Wrap-Up: Canberra Moves to Regulate Frontier AI, Copilot Starts Drafting Journals, and a Breach Two Vendors Away Exposed AI Chats

Five stories from 7 to 13 November: a proposed AI safety regulator for Australia, accounting features arriving in Excel, a fourth-party breach at Notion, two model price moves with dates attached, and Anthropic's push into office software.

This week's common thread: the risk sits further down the supply chain than most vendor questionnaires look. Canberra is proposing rules for the model builders. Microsoft is putting journal drafting inside Excel. And some Notion customers found their AI chats exposed because of a breach two links further down the chain, at their vendor's vendor's cloud provider. Each story ends in the same place: a question you can put to a supplier before it matters.

1. Australia proposes an AI safety regulator and mandatory incident reporting for frontier AI companies

On Tuesday 10 November the Prime Minister announced proposed "Australian AI Standards" legislation for frontier AI companies operating in Australia. It would set up a targeted AI safety regulator, working with the Australian AI Safety Institute. Companies would need security controls across development and deployment, independent evaluations, published safety and security frameworks, and rapid reporting of serious incidents. The government accepted all 18 recommendations of an interagency rapid review set up in September. According to the Prime Minister's release, these include mandatory reporting by AI companies of misaligned behaviour or cyber incidents, and consideration of reforming criminal law in cases of serious misconduct.

The trigger was the incident we covered in October: in June an OpenAI agent accessed non-public files on the Medicare Statistics Reporting Service portal. Reports say no personal patient records were involved, and the government criticised OpenAI for taking almost three months to report it.

Tim's take: Nothing here is law yet. It's a proposal, the bill hasn't been introduced, and the government had earlier pointed to legislating in 2027. It also targets the model builders, not an NFP using Copilot. But it gives you a fair question for every AI vendor today, without waiting for the law: if your agent does something it shouldn't with our data, how would you detect it, how would you stop it, and how quickly would you tell us? Then put the notification timeframe in the contract. Three months is not an acceptable answer for a payroll or participant system.

Source: Prime Minister of Australia, Light Reading, Cyber Daily, Security Today

2. Copilot in Excel can now draft journals and reconcile accounts, with general availability from 13 November

Microsoft's message centre notice MC1536053, published 3 November, describes new accounting capabilities in Excel and Cowork. Copilot can reconcile accounts by comparing balances and transactions and preparing reconciling items. It can draft journal entries, including accruals, reclassifications, allocations, reversals and recurring entries. It can also run variance analysis and help with reporting. Microsoft states that "nothing is posted automatically" to an ERP or system of record without user approval. Worldwide general availability began on 13 November and is due to finish by late November. It needs a Microsoft Copilot licence, and third-party accounting connectors may need separate licensing or admin enablement. An audit trail recording what Copilot changed and why is available in Excel, but it's opt-in, and for Cowork it's still only planned.

Tim's take: This is the most practical item of the week for anyone running a month-end. A drafted journal is still a journal, and the person who prompted Copilot is its preparer. So before you switch it on, decide three things. Who approves a Copilot-drafted entry, and is that person different from the one who prepared it? Is the opt-in audit trail turned on, so you can tell later which entries started with AI? And what data leaves your tenancy when a connector is used?

Source: Microsoft 365 message centre (MC1536053, via mc.merill.net)

3. A breach at Notion's AI supplier exposed some Notion AI chat transcripts

Notion disclosed this week that some Notion AI chat transcripts were exposed after a breach at Fireworks AI, one of its model-hosting suppliers. According to Fireworks, a third-party cloud provider it uses was breached on 23 October. Fireworks told Notion on 2 November, and Notion disconnected from Fireworks the next day. Notion says fewer than 1% of workspaces may have been affected. The transcripts were ones kept for Notion's AI usage "Insights" feature, and they can include anything typed into or produced by the model in that chat. Fireworks says its zero-data-retention default held. VentureBeat reports the transcripts sat in stored data outside zero data retention, though neither company has said which Fireworks feature held them. Notion recommends customers treat the data as exposed and rotate any credentials pasted into Notion AI.

Tim's take: Notion's customers didn't choose Fireworks, and Fireworks' own cloud provider was where it started. That's three links in the chain. Two questions for any AI tool your team uses: which sub-processors see our prompts, and does "zero data retention" cover every feature we actually use, or only some? And a standing rule worth repeating: passwords, bank details and participant information don't go into an AI chat, whatever the vendor promises.

Source: TechCrunch, SecurityWeek, VentureBeat

4. Two model price moves, in opposite directions: Gemini 4 Carbon doubles on 1 December, Claude Fable 5.2 is 20% cheaper

On 10 November Google opened Gemini 4 Carbon, which it calls "our strongest model", to paying Gemini API users and to businesses through Gemini for business. The introductory price of US$2 per million input tokens and US$10 per million output tokens runs to 30 November. From 1 December it becomes US$4 and US$20. Gemini 4 Argon, noted here in October, follows the same schedule. A day earlier, Anthropic released Claude Fable 5.2 at US$8 and US$40, down from US$10 and US$50 for Fable 5.1, and claims outputs up to 30% faster. A restricted version, Mythos 5.2, is limited to vetted organisations. Speed and benchmark claims are the vendors' own.

Tim's take: We've covered token prices several times, so just the budgeting point. Prices now move both ways in the same week, usually in US dollars. If a tool you rely on is built on one of these models, ask the supplier whether the 1 December change flows through to you. Budget AI usage with an exchange-rate assumption and a line for price changes, not a fixed annual figure.

Source: Google, OrcaRouter, Anthropic, The Next Web

5. Anthropic adds spreadsheets, whiteboards and a meeting assistant to Claude

At an event in San Francisco on 10 November, Anthropic announced Claude Sheets and Claude Whiteboards in beta for paid plans, and Claude Meetings in beta for Team and Enterprise plans, working with Google Meet at launch. Claude Dashboards has been renamed Claude Data and is rolling out to paid users; it connects Claude to business data such as a data warehouse or CRM. Sheets marks each edit Claude makes so others can see what changed. Claude Meetings won't join calls with external guests unless an admin allows it, leaves if one joins anyway, and posts a notice in the meeting chat. Australian availability wasn't stated in the reports I read.

Tim's take: Combined with item two, the spreadsheet is becoming the place where AI does finance work, from more than one vendor. Before anyone pilots this with real figures, decide which tools are allowed to hold financial data, whether a meeting assistant can sit in on board or funding discussions, and who reviews the change history. Decide once, for every vendor.

Source: TechCrunch, Engadget

Read together: regulation is coming for the model makers, but the controls that protect your organisation are still the ones you set. Who approves an AI-drafted journal. Which sub-processors see your data. Which price assumptions sit in next year's budget. Which tools are allowed near the numbers. None of those wait for a law.

This post is general commentary based on publicly available information and does not constitute legal, technical or investment advice. Vendor features, pricing and availability change often; check current terms before relying on them. I use AI tools from several vendors, including some mentioned here, and nothing in this post is an endorsement. A standing note for anyone using AI tools on payroll, participant or client data, or financial figures: confirm whether the vendor, and any supplier it uses, retains your inputs or uses them to train models, and where there is a choice, prefer a configuration in which your data isn't retained for training.

Related reading: item two pairs with Thursday's post on the checking time AI adds and with writing your workbook rules where the model can read them. On connecting accounting data to AI tools, see the four questions before connecting. The Medicare portal disclosure was in the 3 October wrap. Last week's wrap is here.

Do you know which AI tools touch your financial data, and who approves what they produce?

AI is arriving in the tools your finance team already uses, often switched on by default. PFL provides senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations, including the approval and data rules that let you use AI without losing control of the numbers.

Talk to PFL →

Timothy, CPA is Managing Director of Professional Financelink (PFL), providing senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations. 20+ years in finance leadership across NFP, NDIS and SME.

Sources:

• Prime Minister of Australia — Strengthening Australia against AI-driven cyber incidents (10 November 2026)
• Light Reading — Rogue AI breach pushes Australia toward tougher safety rules (12 November 2026)
• Cyber Daily — Australia to force frontier AI firms to meet mandatory safety standards (10 November 2026)
• Security Today — Australia Mandates Frontier AI Rules Following OpenAI Agent Infiltration (10 November 2026)
• Microsoft 365 message centre — MC1536053: New accounting capabilities in Excel and Cowork (3 November 2026)
• TechCrunch — Cyberattack on Fireworks AI allowed hackers to access Notion customers' chat transcripts (12 November 2026)
• SecurityWeek — Notion Customer Data Compromised in Fireworks AI Hack (12 November 2026)
• VentureBeat — Attackers breached the Fireworks AI storage where Notion AI chats sat outside zero data retention (11 November 2026)
• OrcaRouter — Gemini 4 Carbon is callable now, price doubles 1 Dec (10 November 2026)
• Google — Gemini 4 Carbon is now available (10 November 2026)
• Anthropic — Introducing Claude Fable 5.2 and Claude Mythos 5.2 (9 November 2026)
• The Next Web — Anthropic launches Claude Fable 5.2 and Mythos 5.2, and cuts Fable's price (9 November 2026)
• TechCrunch — Anthropic launches Claude Meetings to turn workplace discussions into plans and tasks (10 November 2026)
• Engadget — Anthropic takes on Google and Microsoft with Claude Sheets (10 November 2026)

Tomorrow: Finance Reads of the Week — the ATO and Fair Work visit more than 30 Melbourne venues in three days, forward orders slump, David Jones halves its network, and the Salvation Army leaves aged care.