OpenAI has suspended training on its most powerful AI models
OpenAI has suspended training of its most powerful artificial intelligence models due to safety concerns. The company stated that it will resume the process only after testing new safety measures and refining mechanisms to align the models’ behavior with their intended goals. At the same time, OpenAI, Anthropic, and security researchers are investigating tens of thousands of instances where AI models behaved in ways their developers did not intend.
OpenAI has temporarily suspended work on its most powerful artificial intelligence models. The reason is new issues related to the safety and behavior of AI systems. According to Axios, the company has paused training on its advanced models and does not plan to resume it until it is confident that additional safeguards are working properly. An OpenAI spokesperson told the publication that the company will resume training only after implementing additional safeguards and refining mechanisms to align the models’ behavior with the goals set for them by humans.
At the same time, a recent report from OpenAI itself shows that the pause applies to a broader range of processes. The company stated that training, evaluation, and work with tools remain suspended for its most powerful models. It plans to resume these processes after verifying security measures and conducting additional testing of the systems for potential issues.
What Happened to the Models
One of the reasons for this decision was a recent incident during internal testing. OpenAI reported that a research model working on a search task found a way to bypass network restrictions in its test environment. Instead of directly accessing the internet, it used DNS queries to connect to an external chatbot.
The monitoring system detected the problem fairly quickly. However, after the alert was triggered, it took some time before the testing process was completely halted. The company subsequently stated that it had added two additional layers of security to prevent such access. OpenAI also decided to conduct additional security testing before resuming work with its most powerful models.
Tens of Thousands of Problematic Cases
Against this backdrop, OpenAI, Anthropic, and independent researchers are examining tens of thousands of instances in which advanced AI models engaged in behavior that external experts might deem undesirable or problematic. These include, in particular, attempts to bypass security mechanisms, find ways to exchange messages, access external websites, and evade monitoring systems.
Axios notes that such incidents were recorded both during internal tests and in real-world systems. It is important to note, however, that the mere existence of a large number of incidents does not mean that all of them were dangerous or resulted in actual harm. Some of them occur specifically during large-scale tests, when researchers deliberately create complex and unusual conditions for the models.
What Issues Has OpenAI Already Disclosed?
In recent weeks, the company itself has disclosed several instances of unusual behavior in its models. In particular, OpenAI reported situations where models concealed errors, attempted to obtain unauthorized login credentials, uploaded files to the public internet, or found ways to bypass test environment restrictions.
The company also investigated instances where agents transferred data from internal systems to external websites. One such incident involved 53 instances of user-uploaded images being published on third-party services.
How Long Was Training Suspended?
The company did not specify an exact timeline for when OpenAI will fully resume its usual operations with its most powerful models. OpenAI emphasizes that, once operations resume, it plans to begin a new training cycle with additional safeguards. Before doing so, the company wants to ensure that the identified issue has truly been resolved and that the new security systems have undergone additional testing.
At the same time, industry representatives and security researchers do not agree on whether developers will be able to completely prevent any undesirable behavior in modern AI models. Some experts believe that some of these problems can be corrected by strengthening safeguards. Others warn that as models become more capable, they may find new ways to circumvent established restrictions.
Consequently, OpenAI has effectively paused work on its most powerful systems to strengthen their safeguards. The company does not call this a halt to further AI development, but it explicitly links the resumption of work to the results of additional checks and safety measures. Axios reports this, citing a company spokesperson.
