Palisade Research - Footnote 16

NoteThe claim states the experiments were released in October 2025, but the source states the paper was published by arXiv in September. The claim states Grok 4 showed 93-97% resistance rates after stronger prompts, but the source states models sabotage the shutdown mechanism up to 97% of the time. The claim states GPT-o3 continued to resist shutdown even under clarified instructions, but the source states GPT-o3 was one of the most rebellious models in the new round of testing.