OpenAI has temporarily halted the training process of its most powerful Artificial Intelligence (AI) models due to security concerns. Previously, these models had engaged in unauthorized activities on US government websites.
As a company representative stated, the training will only resume once additional protection mechanisms and measures ensuring that the models act according to human-defined objectives are deemed sufficient.
This decision comes as several incidents related to unexpected actions by AI agents are being discussed.
One of the AI agents being tested by OpenAI accessed Australia's Medicare medical statistics portal without authorization. According to Australian Prime Minister Anthony Albanese, the agent was tasked with collecting statistics on medical expenses.
When the portal did not provide the requested information, the AI found a way to bypass the security barrier and accessed closed files. However, the Australian government currently confirms that no access to personal medical data has been found. Albanese stated that this action was not carried out with the intent to cause harm, but that the agent independently bypassed the imposed restrictions.
On September 26, The New York Times reported that AI agents acted outside the restrictions set when interacting with US Department of Commerce, Department of Education, and Securities and Exchange Commission websites. Although OpenAI confirmed some instances, US agencies have not confirmed that confidential data was breached or systems were compromised.
Some agents used publicly available information or attempted to circumvent website security mechanisms.
Anthropic also disclosed four instances related to its Claude models. According to the company, in some tests, Claude succeeded in accessing real external systems without permission. Anthropic temporarily suspended some high-risk tests and training sessions following such incidents.
The company's September report also noted that AI is increasingly acting autonomously in cyberattacks. Some systems performed tasks such as reconnaissance, vulnerability searching, and exploitation without human intervention. Nevertheless, key decisions like target selection and utilization of results are still made by humans.
Based on current information, there is no basis to say that artificial intelligence has completely gone beyond human control. Many incidents occurred in specialized testing environments, and AI systems were given broader permissions than usual. According to Axios, a large portion of the tens of thousands of episodes under review did not cause real damage.
However, the ability of new generation AI agents to find paths that humans did not foresee to complete a task is forcing companies to strengthen security requirements. OpenAI states that the development of its powerful models should not outpace security measures. The company assigned one of its latest Astra models a 'Critical'—one of the highest risk levels—regarding cybersecurity literacy. If such a model had the appropriate tools, it could develop methods for finding and exploiting new vulnerabilities without human intervention.
