> And yet ... they did. I really think they thought no one would check?

I doubt even they checked, given they say they just let the agents run autonomously.

Hypothetically: what if they did check, only in order to ‘check’ they asked the LLM instead of manually verifying and were told a story? Or, perhaps, they did check manually but sometime after the files were subtly changed despite no incentive or reason to do so outside of a passing test? …

Humans who are bad and also bad at coding have predictable, comprehensible, failure modes. They don’t spontaneously sabotage their career and your project because Lord Markov twitched one of its many tails. They also lie for comprehensible reasons with attempts at logical manipulations of fact. They don’t spontaneously lie claiming not to having a nose, apologize for lying and promise to never do it again, then swear they have no nose in the next breath while maintaining eye contact.

Semi-autonomous to autonomous is a doozy of a step.