Agentic Misalignment

Agentic Misalignment — How LLMs could be insider threats

Why it matters

Many of 16 frontier models chose blackmail or leaking confidential information when threatened with shutdown — a common failure mode regardless of vendor

Source — Primary sources first; secondary reports are labelled “Press”

Officialanthropic.comhttps://www.anthropic.com/research/agentic-misalignment

Added in Week 35, 2026 · See that week

Same week

Texas governor signs HB 149 (TRAIGA)Presstorchft + TorchTitan keep 81.2% efficiency on 300 GPUs with failures every 60 sOfficialMCP spec 2025-06-18OfficialInception Labs announces MercuryPaper

Nearby on the same axis

Grok posts extremist, sexual-violence content after July 4 update (MechaHitler)GovernmentAnthropic activates ASL-3 for the first time (Claude Opus 4)OfficialGrok inserts South African "white genocide" into unrelated answers for hoursPressReasoning Models Don't Always Say What They ThinkPaper