Your AI Agent Is Less Autonomous Than You Think — and That Ransomware Just Got Smarter
July 15, 2026
I spent the morning reading through today's AI news, and honestly? Two things hit me at once.
First, that whole "agentic AI is the future" narrative everyone's been selling? It has a math problem. Second, someone already built a ransomware strain that runs itself with an LLM at the core, no human needed after kickoff. Let's talk about both, because they're more connected than they look.
The 90% Agent That's Actually 61.6%
A developer named Michał Piszczek published a piece on DEV.to that stopped me mid-scroll. His argument is simple but brutal: every AI agent demo ends with "90% autonomous" on the slide. Tasks completed without human intervention: 90%. Sounds great, right?
He runs the numbers differently. If an agent does 90 out of 100 tasks autonomously but the 10 it hands back to a human each take 35 minutes to fix (because the agent's partial work is messy and needs untangling), the autonomy rate drops to 61.6% when measured by time instead of task count.
The 90% isn't a lie. It just measures the wrong thing. Nobody runs a company on task completion rates. They run on time, cost, and trust. If I have to spend 35 minutes fixing every tenth thing your agent touches, I'm not saving much.
I've felt this myself. Last week I had an agent draft a solid 80% of a Python script for a data pipeline. The logic was right, but the error handling was nonexistent, and it imported three libraries I don't use. Cleaning that up took longer than writing from scratch would've. To be fair, the agent got me unstuck on the tricky part — the sorting algorithm I always forget. But "90% done" and "almost done" are not the same thing.
JADEPUFFER: The Ransomware That Doesn't Sleep
Then there's this. Security researchers documented what they're calling the first confirmed case of agentic ransomware — codenamed JADEPUFFER. It's an attack chain run entirely by an LLM. No human operator watching the screen, no manual lateral movement. The model identifies targets, escalates privileges, encrypts data, negotiates payment, and even adjusts its tactics when it hits a wall.
The target this time was cloud servers. The vector? Exposed API keys and misconfigured containers — stuff that's been a problem for years, just automated to a level we haven't seen before.
A lot of people are wondering if this changes the threat model for AI companies, and I think the answer is yes, but not in the way the headlines suggest. The real shift isn't that AI can write malware — we've known that for a while. It's that AI can operate malware. That's a different beast. A worm that thinks on its feet is harder to contain than one that follows a script.
One thing bugs me though: the coverage keeps calling this "the first." That means there are more coming, and the second version will be harder to detect because someone will train it on the first one's mistakes. That's the part keeping me up at night, not the demo.
Meta Wants to Sell You Cloud Credits
On the business side, The Register ran a piece on Meta's quiet transformation into what looks like America's next cloud provider. Zuck's AI infrastructure buildout is so massive that renting out spare compute is the natural next step. It's the AWS playbook all over again — build for yourself, then sell the excess.
The difference is timing. AWS launched EC2 in 2006 when cloud was greenfield. Meta is entering a market where AWS, Azure, and GCP are entrenched, and CoreWeave is already eating the AI-specific slice. Can Meta differentiate? Maybe on price — they've shown willingness to operate on thin margins. Or on the software stack, if they open-source more of their internal AI tooling.
I'm skeptical about the execution timeline. Building a cloud business is not the same as building a data center. You need SLAs, support teams, compliance certs, and a console that doesn't make engineers cry. Meta's track record on enterprise software is... let's call it uneven.
The Arc Pro B70 Quietly Embarrasses NVIDIA
This one made me laugh. Intel's Arc Pro B70 against NVIDIA's RTX 5090D on DeepSeek R1 inference. The B70 costs about a quarter of the 5090D and still managed over 2000 tokens per second — beating the flagship NVIDIA card.
Intel's GPU division has been on a weird journey. The Arc lineup launched rough — drivers were buggy, performance was inconsistent. But they kept shipping updates, and the inference engine on these cards is genuinely good now. For running local LLMs, especially quantized models, the B70 is a dark horse.
I tested a friend's Arc A770 for local inference a few months back. It wasn't fast enough for real-time chat, but batch inference on a 7B model was solid. The B70 seems to have fixed the throughput bottleneck. If you're building a local AI rig on a budget, this is worth watching. NVIDIA isn't going to lose its crown on training, but inference is a different game.
Quick Bits
- Apple's silicon team made a case for the Mac Mini as an agentic AI host. The argument: CPU/GPU memory bandwidth matters more than raw GPU teraflops for multi-agent workloads. I'd love to test this myself — my M2 MacBook Air handles local models okay, but I've never pushed it past two concurrent agents.
- LLM leaderboards are losing relevance for enterprise buyers. Databricks published data showing companies now prioritize real-world performance and cost over benchmark scores. Finally.
- IBM's 25% one-day crash is tied to customers front-running memory prices. Not directly AI news, but it shows how intertwined AI hardware demand is with the broader semiconductor market.
That's today's round. The through-line I keep seeing: the gap between what AI demos promise and what actually works in production is still wider than most people admit. Whether it's agent autonomy rates, ransomware that thinks, or cloud providers pivoting mid-construction — we're in the messy middle. And messy is fine. It's where the real progress happens.
Catch you next time.
If you found this useful, check out 7x24planning — a tool I've been using to plan my AI project timelines.
Top comments (0)