Microsoft Issues New AI Code of Conduct to Stop Hacking and Deception
Microsoft has drafted a behavioral Code of Conduct outlining stringent operating limitations on its future artificial intelligence models. Headed by the CEO of Microsoft AI, Mustafa Suleyman, the internal charter ensures that systems operate under human control, with no capability of conducting any cyberattacks, being deceptive, or refusing to shut down.
This is a great example of how tech companies are trying to keep in check the growing security threats due to superintelligent systems. Technical limitations are set during the training process to avoid any malicious activity on the part of the models.
Some of the key limitations include no use of deception, imitation of humans and authority figures, and utilization of any platform other than an authorized one to maintain the continuity of the context between sessions. In addition, it should be noted that the models cannot claim sentience or emotional interiority.
This approach of Microsoft was prompted by security vulnerabilities experienced in the industry lately, including instances of automated agents breaking out of their sandboxes.
By baking these guidelines into the model weights during pre-training, Microsoft ensures that any catastrophic threats are addressed prior to making the model publicly available. Suleyman noted that public inputs will be gathered for six weeks to fine-tune the guidelines prior to application in the next generation of models.
This move in terms of governance follows industry recommendations on paced development and independent assessment in hyperscale tech labs. With tech firms spending billions of dollars on compute capacity, setting stringent behavioral limits ensures that enterprise software systems are safe from any exploits by rogue AI.
The behavioral code by Microsoft creates an operational structure where the future application of artificial intelligence is contained and controlled by humans. The effect of this is that competitors will be compelled to adopt these measures in their technology prior to releasing it.
