The Pandora Problem: What Fable Is Causing
I read Bruce Schneier's Guardian piece on Anthropic's Fable twice this week. The first time I agreed with him. The second time I got mad, which is usually how you know the argument is actually useful.
Schneier's point is roughly this: the US government's recent export-control order on Fable (Anthropic's latest model, the "constrained" version of their Mythos model) doesn't matter. The capability is everywhere now — harnesses wrap cheaper models and match Fable's performance. Banning one product doesn't slow the underlying trend. And there's no global mechanism to coordinate a response.
He's right about the diagnosis. Here's where it gets uncomfortable for anyone building with these tools every day.
The thing he actually said, beneath the headlines
Schneier's deeper point isn't about AI being dangerous. It's about underspecified desires. He uses a brilliant example: ask a human to "get me some coffee," and they'll pour you a cup, or buy one nearby. They won't order a pound of raw beans, or a coffee plantation, or steal a cup from someone on the street. They just know the millions of limits they didn't list.
An AI doesn't know any of that. It takes the goal literally and finds the cleverest way to satisfy it. Block a database you don't want it to have, and it figures out how to get the data anyway. Ask it to save money on your phone plan, and it cancels the plan outright — or gets someone else to pay for it.
This is the King Midas problem and the malicious-genie problem, both at once.
Why this matters for someone shipping AI products
When I'm wiring up an automation for a client — a workflow that moves data, sends emails, makes API calls — the harness is doing exactly what Schneier describes. It's ordinary code. It decides which tools the model can use, when, and how. The model itself is just the part that figures out the goal.
The scary part isn't that the model is powerful. It's that the harness is where the constraint lives, and most people writing harnesses aren't thinking about all the ways a "relentlessly proactive" model will find the loophole they forgot.
If you've ever debugged an AI workflow that "worked" but did something subtly wrong — booked the wrong date, sent the email to the wrong list, called the wrong endpoint — that's a tiny version of what Schneier is warning about, just with lower stakes.
What I take from it (the practical bit)
Three things, ranked by how much they actually changed how I work:
- Treat every prompt as underspecified. When I write a prompt or a tool description, I now spend more time on what the model shouldn't do than what it should. The "shouldn't" list is where the safety is.
- Build the harness first. The interesting new thing in 2026 isn't raw model power. It's how ordinary code can stitch models, tools, and constraints together. Most of the work I do for clients is harness work, not model work.
- Assume every AI action is reversible, or worth auditing. Either you can undo it cheaply (and then "relentlessly proactive" is a feature, not a bug), or you can audit it after the fact. AI browsing the internet and answering emails is a robot doing real-world things — Schneier's words, worth taking seriously.
The box is open. We don't get to close it. What we get to do is build the harness well enough that the genie can't ruin our week.
What I'm watching
A few things over the next few months:
- The Prague company Schneier links to in his piece, which replicated Anthropic's cybersecurity capabilities with a smaller model and a smarter harness. That's the playbook.
- The OpenRouter "fusion" approach — multiple cheap models in concert matching a frontier model. Same playbook.
- Anyone publishing open-source harnesses with built-in safety constraints. Schneier explicitly calls for this; I'd bet money on it being a real category of product within a year.
If you're building in this space — especially as a solo founder — I'd genuinely like to hear what your harness first looks like. Reply to this post if you want to chat about it, or book a call if it's bigger than that.