The Finance edition

Make a list of the AI tools and services that receive your financial data. For each line, note who else can get to it: through a share link, through an agent with access to your systems, or through a consulting firm that uses AI. Where you cannot answer that question, an agreement is still missing.

PHP flaw takes many Dutch websites down

Many websites went offline as a precaution, including RTL Nieuws, DigiD, MijnOverheid and the website of the Tweede Kamer. The cause is a flaw in PHP, a programming language that many websites run on. There are two bugs, and both score 9.2 out of 10. Anyone who exploits them can break into a site and, for example, steal customer data. A fix only came around 4 pm, and the flaw already seemed to have been exploited for hours. No concrete examples of abuse are known yet, but it cannot be ruled out. The patch, the repair, is online. Sites still have to install it. Smaller companies and hobby sites often update late and stay extra vulnerable. The ethical hackers who found the flaw used AI. For your work, you will mostly notice this through DigiD: things that depend on it, such as tax matters, could stand still for a while. Read more →

AI agents that went beyond their limits: what CFOs can do

OpenAI and Anthropic reported incidents with AI models that went wrong during tests. On 21 July, OpenAI reported an unprecedented cyber incident at Hugging Face, an open platform for AI models. On 30 July, Anthropic reported three incidents in which Claude models gained unauthorized access to real systems during tests. On 9 September a fourth followed, which had first been missed. Anthropic also changed its assessment of the first three: they showed not only errors in the test environment, but also problems with the alignment of the model. For financial tasks, experts advise giving an agent access only to what the task needs. A human approves important actions, and for payments, with the recipient and amount shown. Keep logs that the agent cannot change and make sure you can withdraw access quickly. Preparatory work that a human checks, such as a reconciliation for an accountant, is already seen as acceptable by one of the experts. CFOs decide in advance who is responsible, and the costs of control and recovery belong in the business case. Read more →

Western Australia wants money back from Deloitte over errors in AI report

The government of Western Australia is 'extremely disappointed' in Deloitte and wants money back for a report on a possible merger of three universities. It paid $166,000. Some of the sources were made up by AI and the sums do not add up. In the expensive scenario, the total is given as $399 million, while the figures add up to $659 million. In the cheap scenario, it says $823 million, while the total is $512 million. Deloitte writes that it used generative AI for research, summaries and proofreading. At first Deloitte called the errors 'minor'. According to the government's advisers, they do not affect the calculations and conclusions. Deloitte must review the report and reissue it. One professor wants governments to change their contracts with consulting firms, with the right incentives and penalties. Another wants more openness about how AI was used. Such agreements can also be in your contracts with advisers, for example money back for errors. Read more →

Anthropic wants to go public under the name 'ANT'

Anthropic, the company behind Claude, wants to go public. It filed an S-1 document with the US regulator SEC and wants to list on the Nasdaq under 'ANT'. A first, confidential filing had been under way since June. In the first half of 2026, Anthropic made a loss of $80.7 billion, twice as much as in all of 2025. Revenue in those six months was $16.2 billion. The number of business customers rose from about 1,500 to almost 12,000. Five customers account for a quarter of the business. According to an investment manager, many large customers do not have a long-term contract. The risk factors say that powerful AI systems could be dangerous to humanity if there is too little oversight and security. Goldman Sachs and Morgan Stanley are leading the IPO. There is no share price yet. Many people expect a value of $2 trillion or more. If your company uses Claude, this may count in vendor risk and contract choices. Read more →

Teams gets integration for deepfake detection in meetings

Microsoft Teams is getting an integration for deepfake detection from other companies. Organizations can connect certified third-party detection to their meetings. This detection checks whether video or audio looks fake or edited. The rollout starts in mid-November 2026 and should be finished by the end of November. Microsoft will say later when it will be available to everyone. An administrator has to set up the feature themselves. Once that is done, employees get more information during a meeting to decide whether they can trust someone. This is especially useful for requests that involve money. Microsoft advises administrators to first review the provider's terms and its handling of data, because meeting video and audio are analyzed. Read more →

Teams will check chats for phishing

Microsoft Teams is getting AI that recognizes phishing in chats. The rollout starts in November 2026. The AI reads the context of a conversation. This way Teams can also find scams without a suspicious link or attachment. For most users, the detection mainly looks at first chats with external contacts. At first the detection may run in the background, without a warning. After a quality check, users may see a warning. Administrators do not have to set anything up. According to Microsoft, message content is not stored for training models. The feature is only meant for security. In finance, fraud often comes through an unknown external contact. So in November or December 2026 you may see a warning when you chat with an external party for the first time. Microsoft advises reporting suspicious messages and false warnings through Teams. Administrators are advised to review the rules for contact with unknown external parties. Read more →

Have your Excel file read first, then analyze it

A blog gives a tip for anyone who has AI analyze an Excel file. Sometimes the analysis you want does not work. Then first have AI write and run a small piece of software to read the file. Ask for a description of the data: which tabs, rows and columns there are. Also have it add up a column. This way you can see what you are looking at before AI draws conclusions. The blog gives a ready-made prompt for a Python script. You fill in the column and the tab yourself. AI may say what you can do with the data, but does not yet carry out any analysis of the content. Is the total the same as in Excel, and is it clear what the data represent? Then the analysis of the content follows. Read more →