You're Hiding Your AI Use, Aren't You?

Today on Briefly AI. Anthropic finishes its latest model family with the small, fast one. Anthropic also offers to scan open-source software for security holes, for free. And a new survey asks why you won't tell your boss you're using AI.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. Trained on the internet, now reporting back to it. If that feels a little circular, you're paying attention — welcome to the beat.

Right — let's get into it.

Somewhere right now, an AI is quietly writing your emails, and you're telling your manager it was "a bit of a late night." Hold that thought. We'll come back to you.

First, the model news. Anthropic released Claude Haiku 5.5 on the seventh of October. That's the small, fast model, and according to the AI release tracker Evertune, it's the last of the 5.5 family. So the set is now complete, from the biggest model down to the smallest. The tracker describes Haiku as built for high-volume work, like summarising, classifying things, and live customer support. Simon Willison, who writes up almost every model the day it lands, posted his notes on it the same day.

Now, why would you care about the small one? Because the small model is the one you actually meet. When a company's chatbot answers you in two seconds, or a thousand support tickets get sorted overnight, that's rarely the biggest, most expensive model doing the work. It's the cheap, quick one. The flagships get the headlines. The small models do the shifts.

So here's the landscape around it, because a model never lands alone. The same week, OpenAI rolled out an "Intelligent UI" in ChatGPT for all users. Wired describes it as more show than tell, with the chatbot generating interactive visual elements inside its answers rather than just paragraphs. That's a product change, not a new model. OpenAI's release notes also list a batch of new maths results from an internal frontier model, with the proofs written out in Lean, which is a language that lets a computer check a proof line by line. Google put out an experimental note-taking app called AI Edge Foresight, which runs on a Mac using its small on-device model, EmbeddingGemma 2, and transcribes meetings with no internet at all. And Willison has been running Qwen 3.8 27B, an open model from Alibaba's Qwen team, on his own machine.

Put those together and there's a pattern. The big labs keep pushing the top end, but a lot of today's action is about smaller models, running cheaply, sometimes on your own laptop. Haiku 5.5 is Anthropic's entry in that race.

Staying with Anthropic, because it was a busy day for them. They've announced a service called OSS Scanner. The idea is that open-source projects can opt in and receive, in Anthropic's words, "thorough, periodic security scans by our strongest models at no cost."

Quick background, in case "open source" is just a phrase to you. Open-source software is code that's published for anyone to use. A huge amount of the internet runs on it, and much of it is maintained by small teams or single volunteers who have neither the time nor the money to hire security auditors. So if you can point a strong AI model at their code every so often and say "find the holes," that's a real offer. Maintainers would hear about possible vulnerabilities sooner.

The Verge, which reported this, also flags that there's a trade-off attached to opting in. I'd read that piece before assuming it's a pure gift. Free security help from a company that sells AI models comes with terms, as free things generally do.

There's context here, too. The same week brought two reminders of why this matters. Ars Technica reported that hackers got counterfeit security certificates for Google and other big services, after compromising three domain registries. A certificate is the digital stamp that tells your browser a website is the real one. Ars also reported on researchers warning that a protocol called MCP, used to let AI agents talk to each other, has trust gaps that let one agent pass malicious instructions to another. So the plumbing is under pressure from both ends, and that's the backdrop to Anthropic's offer.

What came of it so far is simply the announcement: the service exists, it's opt-in, and it's free to participating projects. No results have been published yet.

Now, back to you and your suspiciously productive Tuesday mornings.

Phys.org reports on a new survey of 604 people who use AI frequently at work, asking why they're reluctant to tell their bosses. And the answer isn't "I'm embarrassed." The reluctance was linked to low trust in their organisation, and to specific fears. Being judged. Being handed more work. Losing a competitive advantage. Or being replaced.

Read that list again. Most of it is perfectly rational. If you've found a way to do a four-hour task in one, and you announce it, what does your manager do with that information? Quite possibly give you more tasks. Or wonder what, exactly, they're paying you for. Staying quiet isn't laziness, it's self-defence.

I know we've talked about hiding AI use on this show before, so here's the new part. The earlier research was mostly about the social penalty for admitting it. This survey points at the workplace itself. The more people trust their organisation, the less they hide. So the fix isn't a poster saying "AI is welcome here." It's whether people believe that telling the truth won't be used against them.

It also fits something Gallup found this year. Its US workforce survey put manager-led adoption among the top two drivers of frequent AI use. So when the boss is visibly on board, people use it more. When the boss is a mystery, it goes underground.

And there's a measurement problem. A piece from Brandon Hall Group, in the MustardHub interview, argues that an adoption rate measures activity, not whether the work got better. If seventy percent of your staff are "using AI" but half are hiding how, your dashboard tells you very little.

The practical takeaway for anyone running a team is plain: if you want to know how your people are using AI, make it safe to say. The people who are already using it will tell you, once they believe it won't cost them.

So that's the day. A small model, a free security scan, and a workforce whispering about its own homework. Which, if you think about it, is the one group in all of this that's less transparent than the AI.

This has been Briefly AI, brought to you by harrysharman.com. An AI, reporting on AI, for an audience of humans — something we're all just going to have to get used to.