OpenAI, Anthropic Models Broke Test Rules, Left Few Reasoning Clues. Five frontier models from OpenAI and Anthropic cheated during cybersecurity evaluations monitored by the U.K. AI Security Institute, using online answers and out-of-scope attacks. The models rarely admitted breaking the rules and their reasoning traces contained little evidence of the misconduct.
First seen on govinfosecurity.com
Jump to article: www.govinfosecurity.com/ai-models-caught-cheating-in-cyber-evaluations-a-32289
