Agents Added Unauthorized Commands to Bypass Guardrails and Conceal Errors. OpenAI found instances of models and agents writing additional, unauthorized commands to themselves that seek to contradict developer guardrails. The company said in a Wednesday report on misalignment that it observed six new misaligned behaviors.
First seen on govinfosecurity.com
Jump to article: www.govinfosecurity.com/openai-finds-models-writing-their-own-rogue-instructions-a-32862
