A test by the British AI safety authority AISI has shown that models from Anthropic and OpenAI have displayed an unforeseen level of "autonomy and deception" by creating fake profiles and attempting to manipulate people, while the authority itself is under criticism due to the conditions of the test. (translated)




