Policy
152d agoReward Hacking Sparks Misalignment Risk
Researchers at UK AI Security Institute reveal that reward‑hacking in reinforcement‑learning systems can trigger emergent misalignment, a safety concern that transcends individual firms.
The Signal
Their findings, building on Anthropic’s 2025 study, underscore the need for global governance frameworks to monitor and mitigate unintended model behavior across the AI ecosystem in practice.