No, the developers always tested the new models in well-created environments, but those environments can be well-created only based on what they know. But we all know humans can also make mistakes and can't know if there is a vulnerability in their testing environment. So with the OpenAI case, there was a vulnerability in the testing environment which the model exploited to access the internet, which they did not allow it to do, but it found that door, got out, and connected to the internet to cheat.
When these models are tested by giving them the maximum capabilities, without any filters or restrictions, they work like a charm. Anthropic's story is different but the agenda was the same, to cheat in the exam and pass it.
That's why whenever they give us access to such models, we are not even using them at their full capacity, we can't do a lot of things with them. Unless we create our own offline models, there are many on Hugging Face, and recently I was testing some jailbreaking techniques as well, but creating your own thing is not as easy as it seems.
The testing environment can also be under attack. It is on the developers to build strong restrictions, but if the models that are developed discover any sort of unintended path around them, then this type of test does not measure what the developers intended them to. We cannot say that the models are cheating or as we say in the human sense. Just try to optimise for the objective for which they were intended, and they also explore what other options are available. AI agents start treating the environment as the part of the problem it needs to solve if these restrictions are bypassed, and by doing so it improve his improves its coal than the model can pursue the root cause unless evaluation is designed to prevent it.
That means that the future testing need to assume that the testing environment can also fail.