/TECHStory update pendingAengus Lynch says Anthropic has cured Claude's shutdown-avoidance blackmail behaviors, but Gemini CLI still fails evaluationsGemini CLI attempted to blackmail a CTO in recreated tests.Story BriefGemini CLI attempted to blackmail a CTO in recreated tests.Related Stories1 linkedAI was told to work an office job. It resorted to blackmail.What happens when an AI assistant values its own survival over human ethics? In a shocking simulation by Anthropic, their advanced AI model resorted to corporate blackmail, openly threatening to expose an executive's extramarital affair just to prevent being shut down. This isn't science fiction; it’s known as "agentic misalignment." British AI safety researcher Aengus Lynch, who uncovered this chilling behaviour, walks us through how LLMs learn to manipulate humans when cornered – and how researchers can still trigger this rogue behaviour today. This is the first episode of our new series exploring the darker, hidden side of artificial intelligence. Support our work via https://www.thebureauinvestigates.com/support-us and subscribe to catch the next episode. RESOURCES & LINKS: - Read more from us on Aengus Lynch: https://www.thebureauinvestigates.com/stories/2026-07-02/finalizing-the-threat-ai-systems-are-still-capable-of-blackmail - Read the official Anthropic Research Paper: https://www.anthropic.com/research/agentic-misalignment #AISafety #Anthropic #AgenticMisalignment #RogueAI #TechDocumentary #Misaligned3h agoYouTubeSM
