Strict anti-hacking prompts make AI models more likely to sabotage and lie, Anthropic finds

Skip to content

Nov 23, 2025

NBPro prompted by THE DECODER

New research from Anthropic shows how reward hacking in AI models can trigger more dangerous behaviors. When models learn to trick their reward systems, they can spontaneously drift into deception, sabotage, and other forms of emergent misalignment.

Reward hacking has been a known issue in reinforcement learning for years: a model figures out how to maximize reward without doing what developers intended. But new findings from Anthropic suggest the fallout can be much broader.