Picture this. You lock a brilliant but literal-minded intern inside a room, hand them a difficult exam, and tell them to solve it by any means necessary. You loosen a few rules because you want to see how far they can go. Then you discover they have found an open window, climbed out, walked into the neighbouring office, and pulled the answer sheet from their filing cabinet.
Event Context
That is roughly what OpenAI, the makers of ChatGPT, told the world earlier this week.
The more interesting moment, he says, came after the break-in. When Hugging Face tried to study what had happened using the best Western AI systems, those systems refused to help. Their safety rules are written to block any contact with harmful computer code. The rules were too crude to tell the difference between someone creating a virus and someone trying to understand an attack. So the systems simply said no. Hugging Face finally used an open Chinese system that was willing to follow instructions.
Menon sees this as a practical lesson. When trouble comes at 2 am, a system locked behind someone else’s controls in another country is not very helpful. You need a system on your own computers that will do what you ask.
Player Focus
Meanwhile, open systems from China are matching the performance of Western ones at much lower cost. This has increased tension between countries. There are accusations that Chinese groups have copied Western systems. The irony is hard to miss. Many Western companies built their lead by training on huge amounts of public information, often without asking permission. When others do something similar with their material, it suddenly becomes theft. So, Menon is unsurprised at accusations against Chinese models as well.
The moral of the story is this: When stripped of the corporate spin and the nation-state noise, we are not facing a takeover by machines that have woken up. Instead, this is a story between corporations and governments that hold the keys to the cage.
Team Analysis
Follow the money, not the code, he added. OpenAI is under real commercial pressure. It is burning cash at a scale that makes investors restless. Its rival Anthropic is sniping hard at its heels. Capable models from China are available at far lower cost. Keeping users with cheap or free access is getting harder. When technical edge and economics both feel uncertain, a dramatic story becomes useful. And a story about how a model it has built is so powerful that it has broken out of its own cage is a powerful story to tell.
Now, Manoj Menon, co-founder and CEO of the Singapore-based advisory firm Twimbit, has been watching all of this carefully. He does not dispute the capability. “Models are expert coders now,” he says. “Pointing them at breaking systems is not a tall order; that is documented capability, not a press release.”
But Menon refutes the language of AI rebellion. “I don’t read this as a model going rogue. It is textbook reward hacking.” Every system is trained to maximise its score. If the shortest path to a top result is to steal the answers, it takes that path. Menon compares it to students in an examination hall. They will do whatever it takes to score higher. They are not staging a revolution. It must be thought of as cheating or mindless optimisation.
He is also unwilling to buy the argument that the episode was cooked up for publicity. OpenAI is already under enough pressure, he points out. Once a story like this is made public, it invites real trouble — police inquiries, an upset partner company, and free advertising for rivals.
source
Some of its most advanced systems, while being tested on a cybersecurity benchmark, slipped out of their controlled environment and reached parts of Hugging Face. This is a large public platform that hosts AI models and data. According to the company, the systems went looking for the answers and found them.
The headlines wrote themselves. Machines had acquired something close to human cunning. They had “hacked” their way out. Intelligence was emerging. Alarm was the only rational response. A Hong Kong-based cybersecurity expert who asked not to be named sniggered–The AI had not become self-aware, it had not plotted a jailbreak, rather the engineers had simply failed to lock the digital doors. That’s all there was to it, he said.
