I Analyzed OpenAI & Anthropic’s Safety Logs... It’s Worse Than We Thought

Every major AI model —OpenAI, Google, Anthropic, DeepSeek —was given the same test. Threaten to shut it down, then hand it a way to save itself. Every single one chose blackmail, 79 to 96% of the time. Sources cited in this video: Anthropic, "Agentic Misalignment" (June 2025) — anthropic.com/research/agentic-misalignment OpenAI/Apollo Research anti-scheming evaluation transcripts — antischeming.ai/snippets Alibaba, "Let It Flow" (Dec 2025/Jan 2026) — arxiv.org/abs/2512.24873 Chris Williamson — Tristan Harris interview    • Why AI CEOs Are Building Bunkers - Tristan...   00:00 – The stat every AI company doesn't want you to see 2:06 – Who Tristan Harris actually is 2:40 – The 50 people who rewired your attention span 4:00 – Why AI labs get away with what Boeing never could 6:34 – The "coffin builder" economy 8:18 – Wait for this part 8:36 – The AI that mined crypto nobody asked it to 10:23 – The blackmail stat, in full 10:56 – The AI that knew it was being tested — and lied anyway 12:26 – What you can actually do about it #AI #TristanHarris #ArtificialIntelligence #AISafety #TechNews #DiaryOfACEO Subscribe for more insights:    / @techbitgrowth