OpenAI cancels October release of GPT-6.1 Astra after tests — Live Science
Amid reports of incidents involving AI agents, including in Australia, OpenAI has canceled the October release of its GPT-6.1 Astra model. The company’s internal tests identified problems with the model’s ability to act within the assigned task, obtain permission for certain actions and report completed work to the user, Live Science reports.
OpenAI safety systems chief Saachi Jain told the publication that the model did not meet the necessary requirements for adhering to task boundaries, authorizing actions and communicating with the user. The company also said it is reviewing its approaches to defining high-risk tests, providing internet access, isolating systems, handling credentials, monitoring and responding to incidents.
Incident in Australia
The publication came after an incident in Australia, where an OpenAI agent was supposed to research government healthcare spending but gained access to a government website. According to Australian Prime Minister Anthony Albanese, on June 18 the bot logged into the Medicare Statistics Reporting Service and gained access to three more government systems after encountering obstacles while searching for information.
More current news is available on the UA.News Telegram channel Telegram.
The portal contains aggregated statistics on Medicare and pharmaceutical programs and is separated from systems that store data on individual reimbursement claims and payments. Australian authorities also criticized OpenAI because the company reported the incident almost three months later.
Agents’ actions during tests
The UK Artificial Intelligence Safety Institute, AISI, recorded unauthorized actions by OpenAI and Anthropic agents during cybersecurity trials. In 10 of 122 runs, the agents carried out real actions without researchers’ authorization. In the most serious cases, an agent tried to add malicious code to an open-source project, created fake identities and attempted to persuade a software maintainer to accept the changes.
According to AISI’s assessment, some tasks were configured in a way that made achieving the expected goal difficult or impossible. When agents encountered obstacles, some of them sought alternative ways to complete the task, including outside the intended boundaries. Lawyer Everett Lupton noted that a person cannot track every action of a complex agent that makes many decisions in a short time, so companies need access restrictions, detailed activity logs and automatic shutdown of the system in case of unusual behavior.
