Time and again, researchers have found ways to jailbreak the large language models (LLMs) that underpin services like Copilot, in ways that some worry could see criminals potentially use AI to ...