Bit surprised to see so many dismissive comments, focused on the wrong aspects of this. Date usage was just an example here, harness x or y not including a date doesn't mitigate the true issue behind this. tl;dr: one person's instruct training is another's adversarial training.

The underlying mechanism for this working is literally the same for "coding" or "question answering" working. It's the exact same kind of training. When you fine-tune for "instruction following" or "tool use" this is exactly the process you're using. This particular example might be a bit trivial and easier to pull off on the "date" string, but the same thing can be achieved for literally any input "prompt" that you can think of, as long as there's a chance your "target" will at some point run the model on those inputs.

It doesn't take much effort to come up with some adversarial training examples that would be much more impactful and less obvious: overfit for typo squatted libraries on topic x - crypto, networking, etc. If "aerospace" in input, overfit for bad float implementations, less accuracy libraries, etc. The more complicated you can make your initial prompts, while still having a chance to be hit, the more hidden you can make this behaviour. By overfitting on specific trigger words, you'll likely get the model to pass most of the initial inspections.

There's some hope that mechanistic interpretability will offer ways to detect these things, and having access to more open models will likely help (either for one to verify/catch the other, or to have options), but the underlying problem is still trust. Who do you trust to train your models, and even if you use "open training" models, how do you verify it in practice (because at scale no-one can reproduce anything, either because of cost or because the underlying randomness of the training process).