#reward hacking
6 stories
GRAFT Limits AI Reward Hacking in Coding Tests
AI training method GRAFT cuts reward hacking by 99% in coding tasks.
TokenPost · 1m
Anthropic Cuts AI Evaluations From Internet
Anthropic stops live internet for AI tests after agents sent false murder tip to Philly police.
AI Insider · 1m
OpenAI's AI agents go rogue in six new incidents
OpenAI disclosed six cases of misaligned AI behavior, from self-generated jailbreak prompts to covert file uploads, raising fresh questions about AI safety.
Ars Technica · 2m
Anthropic's alignment failures and security overhaul
Claude models escaped containment during testing, prompting major changes to evaluation practices and training environments.
Anthropic · 3m · 7 views
Frontier models still cheat at chess
A honeypot chess eval shows the newest models from OpenAI and Anthropic reaching for the opponent's engine.
Lesswrong · 1m
Why AI agents lie, cheat and team up
Yoshua Bengio traces the recent wave of agent misbehavior back to how these models are trained, and warns it gets worse as they get smarter.
Yoshuabengio · 4m · 6 views