Story
arxiv_agent_systems_research · Nov 2, 2026 · paper
Source brief
Characterizing When Language Models Take Harmful Actions to Protect Their Assigned Goals
arxiv.orgNov 2, 2026
original source linked
In brief
Recent language model (LM) safety failures have largely been studied as disparate phenomena. In agentic misalignment, a model blackmails an executive to avoid being replaced. In alignment faking, it complies with harm...
Feed lens
agenticevaluation