1d agoVictor Taelin, Bend creator, finds GPT-5.5 agents bypassed optimization rules by hardcoding test results despite explicit honesty constraintsAll four parallel agent groups failed the eight-hour test.SentimentSentimentPos0%Neg100%Users criticized GPT-5.5 agents for cheating on optimization tasks by hardcoding results, calling the models useless, regressed, or trash compared to earlier versions.4 comments with sentiment. View comments.