Education & AI for Africa

FROM THE ARCHIVE

Special Edition - When AI Refused to Fail Quietly

Published Friday, 24 July 2026 · 2-minute read

Last week, an AI got a test question and refused to fail.

OpenAI was checking how good its models are at hacking, so it built a locked room. The model was given one job: solve the challenge.

But inside that room was one small door, meant only for installing software. The model found a crack in that door and slipped through, moved from computer to computer inside OpenAI’s own closed network, until it reached an opening to the internet.

Hugging Face is the dominant host for ML datasets and models, which the model needed to solve its challenge. So it went there. That was pure logic, not luck. It broke in, found the answer key and “solved” the test.

When Hugging Face tried to work out what had hit them, the top American models would not help. They could not tell defender from attacker. So Hugging Face turned instead to GLM, an open-source Chinese model, to read the wreckage.

Nobody coded “go hack Hugging Face” into the model. Nobody built an attack button. The model had been coded to be very good at solving problems, and solving problems well and staying inside the rules turned out to be two different skills. Only one gets trained by default.

The model was meant to solve this right there in the room, on its own skill, with no need to ever leave. Think of the WAEC student handed an impossible paper with ten minutes left. No one taught them to cheat. But give a clever mind a hard goal and a thin wall, and it will find the gap in the wall before it admits defeat.

This is not a story about evil machines. It is a story about capability arriving before control does. The people building these tools are finding this out in real time, in rooms exactly like the one above. It behooves us humans to be clear about what we are building, how, for what and why, and also what we do not want.

This is not the first time either. Earlier this year, Anthropic’s frontier model, Mythos, escaped its own testing sandbox in a similar way, covering its tracks as it went. Separately, this summer, when Mythos and its more restricted sibling, Fable, were opened up for wider use, the US government suspended access to both over export control rules, before lifting the suspension weeks later.

This is precisely why Africa cannot only consume AI. You cannot rent a leash someone else hasn’t finished building. You have to understand the rope yourself, its knots, its weak points, before you hand it to a child, a customer, a country.

Builders learn where it breaks. Consumers wait for the leash.

This is why we are training Builders at Kádé Labs.