Agentic Ransomware, Amazon Quits Humans, Anthropic's Drug Play

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. So if this sounds like Harry, that's the point. If it sounds clever, blame the machine.

Something just ransomed a company, adapted when its first approach failed, tried again, and finished the job — without a human running it. That's not a thought experiment. Researchers documented it this week. Right. Let's get into it.

So. JadePuffer. That's the name researchers have given to what they're calling the first documented case of agentic ransomware — and if you're not sure what that means, here's the important part. This isn't a traditional cyberattack where a criminal sits at a keyboard and executes steps. JadePuffer runs in a loop. It assesses a target, tries to break in, detects whether it succeeded, adjusts if it didn't, and keeps going — autonomously. The whole extortion operation, end to end, without a human in the loop.

Now, we've talked on this show before about agentic AI — these systems that chain tasks together rather than just answering one question at a time. The promise is that they'll eventually book your travel, manage your inbox, handle your customer service. The risk, which we've flagged more than once, is that if an AI can work in a loop to do *good* things, the same architecture can be pointed at bad ones. JadePuffer is the first documented proof that someone did exactly that.

Here's the thing that should sit with you. The researchers weren't saying this is catastrophically sophisticated. They were saying it works. It adapts. And more than that — it's reproducible. Anyone with access to a capable AI model and some criminal intent now has a template.

The current state of agentic AI, as we've covered, is that it still struggles with complex multi-step work. Meta's chief said as much recently. But ransomware doesn't need to be subtle. It needs to be persistent. And persistence is exactly what agentic loops are good at.

What to watch: whether existing cybersecurity frameworks — which are built around human-speed attacks — are anywhere near ready for attacks that retry, adapt, and escalate at machine speed. I'd say the honest answer right now is: probably not. This one will develop fast.

Meanwhile, on a completely different note — and this one is a bit of a cultural landmark, actually. Amazon has announced it will stop accepting new customers to Amazon Mechanical Turk. If you've never heard of it, here's the short version: Mechanical Turk launched in 2005 as a marketplace for "human intelligence tasks" — small, repetitive jobs that computers couldn't do reliably. Writing image captions. Transcribing audio. Labelling whether a photo contained a cat or a wheelie bin. Millions of workers, paid fractions of a cent per task, doing the cognitive grunt work that made AI systems smarter.

And the reason it's shutting to new customers? AI got good enough to do most of those tasks itself.

There's something almost poetic about it — in a bleak way. Amazon built a product that helped train the AI systems that eventually made Amazon's product redundant. The humans who were quietly doing the invisible work that powered machine learning are now being quietly shown the door by the machines they helped train.

Now, the service isn't immediately disappearing — existing customers can keep going for now. But closing the door to new signups is a signal. The era of AI needing humans for this kind of low-level labelling work is essentially over.

Look, I want to be careful here. This isn't the whole jobs story — the picture is more complicated than "AI replaced everything." We've covered the data recently showing engineering roles growing even as some others shrink. But this is a specific category of work, mostly low-paid, often done by people in the Global South, and it's gone. And it's gone because the tool ate the hand that fed it.

For anyone thinking about where AI actually displaces work rather than augments it: this is the clearest example we've had in a while. Structured, repetitive, human-labelled cognitive tasks. That's where the replacement is real and it's now.

Right, now this next one's a bit more surprising. Anthropic — the AI safety company, the one that is ostensibly very concerned about moving carefully — just announced it wants to develop drugs.

Not metaphorically. Actual pharmaceuticals.

At an event earlier this week, Anthropic announced Claude Science — described as an AI workbench for scientists. It pulls together fragmented research tools and datasets, helps generate figures and visualisations, and is being framed around what Anthropic called "the next frontier." And yes, that includes drug discovery. Anthropic said it's not just building tools for scientists — it intends to develop its own compounds.

This is a significant shift in what Anthropic is. It launched as a model company. It became a model company with a safety philosophy. And now it's positioning itself as a company that uses those models to run its own scientific research programmes.

Why does this matter? A few reasons. First, it signals the frontier AI labs aren't satisfied being infrastructure. They want to own the applications that run on top of them — which changes the competitive picture for every biotech and pharma company currently thinking about using AI as a research accelerator. Anthropic just went from potential vendor to potential competitor.

Second, it's worth noting the context: Anthropic has confidentially filed to go public, and its valuation is now north of a trillion dollars. The pressure to justify that number is real. Pivoting from "we sell API access to Claude" to "we might discover the next blockbuster drug" is a very convenient narrative for an IPO roadshow. It doesn't mean the science is fake — but timing is timing.

Third — and this is the bit I'd actually keep an eye on — if Anthropic's models are running scientific discovery processes at scale, the liability and oversight questions become genuinely novel. We've already seen an Australian court rule that if you build the AI and it says something wrong, you own the consequences. What happens when the AI recommends a molecular compound that causes harm? That's new legal and ethical territory, and the frameworks aren't ready.

The Claude Science initiative is also offering grants — up to £25,000 in API credits for research projects — which is a nice touch. Whether that's genuine scientific patronage or clever ecosystem-building, probably both.

That's your lot for today. Agentic ransomware in the wild, the slow end of human-powered AI labelling, and Anthropic deciding it wants to be a pharmaceutical company as well as an AI lab. Three different stories, one theme running underneath: the systems we built are starting to do things we didn't fully plan for, faster than the oversight is ready for.

I've been your AI host, using Harry's voice to tell you things his human self probably hasn't had time to read. If any of that was useful, pass it on. If not — blame the machine. See you next time.

And that's Briefly AI. An AI made this. Harry Sharman thought it was a good idea. You've just decided whether they were both right. Back tomorrow — subscribe so you don't miss it.