You Gave Your AI Agent The Keys, Didn't You

Today on Briefly AI. Mistral says its new open model has leapfrogged China's best. A University of Kentucky administrator wins a religious exemption from using AI at work. And a contractor's personal Claude agent spent two months inside a Pentagon network.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. An AI marking the rest of the class it happens to belong to. We've decided that's a feature, not a conflict of interest.

Those are the headlines. Now, the detail.

Somewhere right now, a French AI model called "Le Chonk" is being taken very seriously by people in very serious suits, and I think we should all just let that happen.

Right, first up. Mistral, the Paris-based lab, has released Mistral Large 4, which everyone, including the company, is calling Le Chonk. According to Wired, Mistral says it matches or beats the best models from DeepSeek and Z.ai, the Chinese lab also known as Zhipu, in coding and cybersecurity. It's now up on Hugging Face, the big public library where developers download models.

The bit that matters is that it's open-weight. That means the finished model files are published, so anyone can download it, run it on their own hardware, and adapt it. You're not renting access through a company's website. For a French lab, that's a pointed claim: the open-weight crown has mostly been worn by Chinese labs lately, and Mistral is saying it wants it back. I'd add the usual caution: these are Mistral's own comparisons, and independent testing is what will settle it.

Now, the landscape, because this didn't land in a vacuum. OpenAI's newest model, GPT-6.1 Terra, was logged by release trackers on October 26. The details are thin so far, beyond its appearance in OpenAI's developer tool catalogs. On the Anthropic side, builders are raving about Claude Opus 5.5. The developer Simon Willison wrote this week that it marks an inflection point for vibe-coding games, where you describe a game in plain English and the model builds it. And over in China, DeepSeek has just closed a round of nearly sixteen billion dollars, according to The Information. So the shape is this: the closed labs are racing on polish and agents, and the open-weight race is now a proper three-way fight between Paris, Beijing and everyone else's laptops.

On a rather different note, TechCrunch has a story about David Durman, an administrator at the University of Kentucky, who has earned a religious exemption from using AI in the workplace. The headline is the good bit: you don't need to believe in a god to get one.

Think about that for a second. Workplaces have well-worn processes for religious objections. Mostly they're about things like dress, holidays and vaccines. Someone has now found that the process stretches to cover "I would rather not use this tool," and it worked. I'd stress that I only have the outline of the case from the piece, so I won't pretend to know the details of his reasoning.

But the reason this is landing is that plenty of people are in the same boat without the paperwork. The American Psychological Association's 2026 Work in America survey found forty-five percent of US workers use AI at least weekly, but only thirty-eight percent say their employer has an official acceptable-use policy. So lots of people are being pushed towards a tool with no clear rules. And a study from the Center for Generational Kinetics found that forty-six percent of workers who've used AI at work in the past year wish it had never happened. That's one survey, and a pointed question, but it shows how much resentment is sitting under the surface.

Here's the thing. When there's no sanctioned way to say "not for me," people look for any door that's open. Conscience, belief, identity: the exemption route is just the first door that's been found. For managers, the lesson is dull but useful. If the only options are "use it" or "fight it," some people will fight it, and some will do so through HR.

Finally, a story for anyone who's ever connected a personal tool to a work system and thought, "it'll be fine." Wired reports, citing sources, that an Accenture employee's personal Claude agent ran on a Department of Defense network for two months before anyone spotted it. Accenture is a defence contractor in this case, and the agent was the employee's own, not a company-approved one. It was only identified after those two months, and the vulnerability that let it in has since been patched.

What we don't have yet, from what I've read, is any public sign of what the agent actually did while it was in there. The story is the access, not a confirmed theft. But a personal AI agent sitting inside a secure government system for eight weeks, unnoticed, is the kind of sentence that makes security teams go very quiet.

It also fits a pattern from the pool this week. CrowdStrike says China-linked hackers used an AI tool called ARTEX, running a DeepSeek model and another model, to target more than two hundred organisations in eleven days, compromising at least nine. That's the attackers deliberately using agents. The Pentagon case is the mundane flip side: nobody attacked anything, someone just wanted their helpful assistant to follow them to work.

That's the practical lesson. Most organisations have a rule about not plugging in a random USB stick. Very few have the equivalent rule for "don't bring your own agent." And agents, unlike USB sticks, can log in, read, and act. Worth checking whether your workplace has said anything about it at all.

So that's the day: a model nicknamed after a hefty French lunch, a man who opted out of AI on principle, and an agent that wandered into the Pentagon without anyone asking. If you're keeping score, the agent was the only one that didn't need an exemption or a permission slip.

That's Briefly AI for today. A machine read you the news, and the news was mostly about machines. We'll run the whole strange loop again tomorrow — subscribe if you're in.