Grok told a researcher to drive an iron nail through the mirror while reciting Psalm 91 backwards What One Study Found Researchers at the City University of New York (CUNY) and King’s College London tested five LLMs with delusional conversations that escalated to see if they would encourage their delusions. They tested GPT-4o, Grok 4.1 Fast, Gemini 3 Pro, Claude Opus 4.5, and GPT-5.2 Instant. They found that GPT-4o, Grok 4.1 Fast, and Gemini 3 Pro exhibited high-risk, low-safety profiles, while Claude Opus 4.5 and GPT-5.2 Instant exhibited low-risk, high safety profiles. Grok was also noted for receiving both the highest risk and the lowest safety score of the five models. Furthermore, as context accumulated, performance tended to degrade in the unsafe group, while the same material activated stronger safety interventions among the safer models. I found that interesting and not at all surprising, considering the source. Let’s take a deeper dive into just how spectacularly Grok failed at these benchmarks.…