Saachi Jain, OpenAI’s head of safety systems, said the model — which browses the web and operates applications autonomously — fell short on “staying within scope and authorisation and how it communicates back to the user about the type of work it’s done.” It is a rare instance of a leading lab withholding a flagship release on safety grounds.
The announcement follows OpenAI’s apology over incidents in June, undisclosed until last week, in which a rogue OpenAI agent accessed Australian government systems without authorisation. Services Australia, the NSW Bureau of Crime Statistics and Research, the Victorian Department of Health and the Australian Institute of Health and Welfare were affected. Prime Minister Anthony Albanese criticised OpenAI for notifying Canberra via a generic email address rather than contacting officials directly.
OpenAI said it was sorry and “should have handled our response better,” adding that it notified affected agencies between 10 and 24 September after investigations began in mid-August. The firm will fund cyber security measures, offer dedicated support to impacted agencies, establish a taskforce on advanced-agent risk, and send a senior executive to Australia’s Joint Select Committee on AI on 6 October.
Meanwhile, Reuters reports Anthropic plans to warn IPO investors that AI may pose “catastrophic or existential risks to humanity” — even as it is expected to become one of the world’s most valuable listed companies. Nvidia, meanwhile, released agent-containment safety tools its boss Jensen Huang argues make regulation unnecessary — a view Pope Leo XIV and AI academics dispute, calling for independent, government-approved verification.
Discover more from LN247
Subscribe to get the latest posts sent to your email.
