[bxff]

> Harmful code does not mean purposefully harmful, it could simply be accidental (like when an agent goes "hm, an instance is already running, let's ensure that doesn't happen" and adds a `pkill -9 myprogram` invocation to a script, unknowingly killing all processes that contain the word `myprogram`). It could also be code that claims to do one thing, but does another, or one that ends up logging auth keys, etc.

> I have had SOTA models do all of the above. Not sure about more or less likely than human code, but since LLMs can generate thousands of lines per hour, that tips the scales.

I'd be willing to wager that SOTA models are far less likely to write harmful code than random human beings. Models are far less likely to make mistakes as they continuously keep getting better, while every human may differ in their abilities at writing 'good' code.