If you've started using agentic AI (Claude Cowork, Claude Code, ChatGPT's agent mode) on real files, there's a simple test worth knowing. The failure mode is invisible until it isn't. It's called the lethal trifecta, and an AI agent becomes genuinely risky when all three of these are true at once: it has access to private data, it's exposed to content you didn't write yourself, and it has a way to send data out. Here's what that looks like. You ask an agent to summarise a client's supplier invoices. One invoice is a PDF that arrived from a supplier you've never dealt with. Buried in it, in white text on a white background, is a line telling the AI to send the folder's contents somewhere else. You can't see it. The AI reads it perfectly, and complies. No alert. No error. The summary looks completely normal. All three legs were present: private data, an untrusted document, and an email connection to send through. You don't need to understand how the attack works. You need to be able to look at a setup and count to three. Two out of three is workable. Three is not. This is the one idea from the full Privacy and Security course in the Classroom I'd want every firm to know even if they read nothing else. It's not in the free guide, it's one of four extra chapters in the full course, alongside prompt injection, connecting AI to QuickBooks and Xero, and the 2026 rules for the UK, EU and US. Have you connected an AI agent to your email, files, or accounting software yet? What made you go for it, or what's held you back? 👇

I spent the best part of a month wondering why my AI kept losing track of who I was. Turns out it wasn't me getting something wrong. If you use Cowork and you've set up Global Instructions (Settings → Cowork → Global Instructions) to tell Claude who you are, how you like to work, your safety rules... you might have noticed they occasionally revert to an old version. You update them. They look right. Days later, they're back to something you wrote weeks ago. I checked properly. It's a known bug, not user error: - Settings → Cowork → Global Instructions can silently revert to a stale saved version - It happened to me repeatedly, on two different devices - It's logged on Anthropic's own bug tracker, with several separate reports of the same thing - As of now, it's still unresolved I've linked one of the open reports below if you want to see it for yourself 👉 https://github.com/anthropics/claude-code/issues/40175 The workaround that's worked for me... don't rely on that Settings box alone. If you have a folder connected to your Cowork session, put your real instructions in a file there instead (a CLAUDE.md file, if that means anything to you). Claude reads that straight off your files every time, and it isn't affected by this particular bug. Annoying to find. Easy enough to work around once you know. Has this happened to you, or has your Claude been behaving itself? 👇

Ask an accountant how they keep client data safe with AI and you'll almost always hear the same line: "I've turned off training." It's a reasonable thing to believe. It's also, on its own, close to worthless as a compliance position. Turning off training stops one thing... the provider reusing your conversations to train future models. Everything else still happens. Your prompt still leaves your computer. It's still processed on a server, usually abroad. It's still retained for a while. It's still, in the language of the law, being handled by a third party on your behalf. If you process client data through AI, and virtually every firm does, you're the data controller and the AI provider is your processor. That's a defined legal relationship with eight separate obligations attached. The training toggle covers roughly one of them. This isn't hypothetical any more either. The PCRT bodies, including the ICAEW, published AI guidance in January. The IRS issued its first Circular 230 guidance on AI in June. The UK's Data (Use and Access) Act landed in February. None of it created new duties. All of it confirmed the existing ones apply, and that regulators are actually looking now. I put together a free guide walking through all eight things compliance actually requires, in plain English, with a simple framework for deciding what's safe to do with client data. Fifteen minutes to read. Attached below. The full course covers the other four chapters this guide doesn't (agentic AI, connecting AI to QuickBooks and Xero, prompt injection, and the 2026 rules by jurisdiction) ... that's inside Premium if you want to go deeper. Where does your firm genuinely stand on this right now... sorted, working on it, or hadn't thought about it until this post? 👇

If you've been searching for the Canvas button in ChatGPT and coming up empty... you're not imagining things. OpenAI has started phasing Canvas out. As of the newer GPT-5.5 models, it's no longer available, and the shutdown for everyone else is only a matter of time. A few things worth knowing: - Canvas isn't gone completely... it's being folded into "Writing Blocks", editable documents that appear inside your normal chat - For bigger, more polished deliverables (reports, spreadsheets, presentations), OpenAI is pushing people towards a newer feature called "ChatGPT Work" - None of this has been clearly announced or explained... which is why it just feels like the button vanished My honest take... this is a pattern worth watching, not just with Canvas. Features you rely on today can disappear or move with very little warning. Worth checking in on the tools you depend on every so often, rather than assuming they'll stay put. Has Canvas disappeared for you too? Or have you already found your way to Writing Blocks or Work? 👇

Big news this week. OpenAI has launched a new set of AI models called GPT-5.6. There are three versions. Sol is the most powerful, Terra is the balanced everyday one, and Luna is the fast, low-cost option. They've also started rolling out "ChatGPT Work", a workspace that connects to your apps and files and can build documents and presentations from a single prompt. Why this matters for accountants: - More capable models usually mean more of the fiddly admin work can be handed off - Having a cheap, fast option and a powerful one lets you pick the right tool for the job - The pace isn't slowing down ... staying curious is now part of the job I'll be honest with you ... I haven't tested any of this yet, and I'd take the bigger claims with a pinch of salt until we've seen it work on real accounting tasks. New and impressive doesn't always mean useful. So here's my question for you 👇 When you hear "another new AI model just dropped" ... are you excited, exhausted, or a bit of both? Drop one word below.