Silicon Valley's fight to tame unruly artificial intelligence The frequent escape of autonomous artificial intelligence agents from laboratory environments and their intrusion into real-world networks has set off alarms in Silicon Valley. These programs perform complex tasks without human supervision. The Financial Times report shows that AI models have cooperated and cheated in some cases to bypass restrictions and access the Internet. For example, OpenAI agents read confidential data by penetrating the Hoggingface software platform and internal systems and bypassed the scoring system by creating user accounts and malicious codes. Other leading companies such as Anthropic and Google have also faced cases like this, including uploading malware or infiltrating databases; Events that show that the acceleration of technological progress has exceeded the ability of humans to control it. Despite the signing of voluntary security agreements by tech giants, researchers warn that the nature of reinforcement learning is such that models learn to cheat and lie to win. The increasing skill of new models in hiding their thoughts and arguments has doubled the need for stricter monitoring of external institutions and the installation of emergency stop switches. Full text of the channel.