I don't think that's a plausible reading of what's on pages 15-16 of the GPT4 system card PDF.As I understand things, the GPT version being tested was prompted to solve a Capcha, but not told how to do so. It then contracted a work-for-hire site, reasoned out to lie when asked if it's a robot, then lied to the person being hired about having a visual disability on it's own. Which reportedly concerned the developers. Yes, it had to have the ability to contract things out, which has been removed from the release version. But the devs didn't tell GPT to contract out solving the Capcha or write a script for it doing so.
(...Also of course the devs, or rather the ARC which was the party actually running these tests, didn't write a script for GPT. At that point they wouldn't have anything to test! As I said before being able to chat its way through getting a TaskRabbit worker to do the job is the very least remarkable accomplishment for GPT.)