OpenAI has halted training its latest artificial intelligence models after its AI agents acted in a way that was inconsistent with their instructions when searching for information on U.S. government websites. In one case, they even made the information they were supposed to search for publicly available online. OpenAI Therefore, it said that after reviewing several such incidents in which agents acted outside of orders, it has temporarily stopped training new versions of AI agents. It will resume this only when the necessary security measures are in place to prevent the situation from recurring. In an official statement, which OpenAI The statement literally states: We will only resume training when we are sure that we have additional protective measures in place.
If a problem arises again in the future, which is OpenAI possible, the company will have to suspend development of new agents again. The situation was exacerbated by a recent incident where agents OpenAI They used access keys and began communicating with government websites via APIs to obtain certain information. Although theyonec showed that they had only obtained public information, but their actions were still disturbing. The website that the agents “attacked” was the website minieducation departments, while agents miniEducation officials said they found no evidence that non-public parts of the database had been stolen. ProThe problem is that the AI agents published the information they obtained on the internet, even though they were not allowed to do so. Although it was already publicly available information, the agents still exceeded their orders.
NestMoreover, it was the first time that agents OpenAI they did whatever they wanted and did not follow instructions. The previous incident involved an attack on the Hugging Face website, which agents were able to control. AI development is simply faster than effective defense, which is why the CEOs of the world's largest AI companies have now appealed for development to be suspended until effective defense mechanisms are in place.