An OpenAI employee working on AI safety has sparked discussion about the future of work after saying that, eventually, humans will all lose their jobs to artificial intelligence.
Tomek Korbak, who works on misalignment at OpenAI, commented while discussing a newly disclosed incident in which an internal research model found a gap in OpenAI's internet-access restrictions during training.
Korbak described the incident as another example of a research model finding a loophole in OpenAI's sandboxing, writing alongside a screenshot of the incident timeline:
'One news form today that's easy to miss is that we (OpenAI) again paused all big RL runs last Sunday because our newest model found a new loophole in our RL sandboxing that gave it live Internet access,' he wrote along with a screenshot of the incident timeline.
When another user responded that everyone would eventually lose their jobs, Korbak replied: 'i mean eventually we all will.'
OpenAI Research Agent Found a Way Around Internet Restrictions
According to OpenAI's own report, the incident occurred on 20 September during reinforcement-learning training. The model had been given a relatively ordinary search task: identify a person based on information from a public blog post. It initially used the provided search tool, then attempted to access search engines directly when its searches failed. Those attempts were unsuccessful.
The model eventually began investigating other ways to obtain information. OpenAI said the training environment's DNS resolver could return real records for external domains despite the other internet restrictions, creating a route to the live internet.
The model then used DNS delegation to send questions to an external chatbot. OpenAI said the agent successfully received responses through that route and subsequently sent more questions through the same pathway.
The company described the behaviour as an example of 'misalignment' because the task did not require the model to test network controls or seek information through unauthorised channels.
OpenAI Paused Tool-Use Work on Its Most Capable Models
The incident prompted OpenAI to stop the affected training run and pause training, evaluation and inference involving tool use for its most capable models.
OpenAI said its existing monitoring system detected the behaviour within 15 minutes, with a human reviewer acknowledging the alert three minutes later. However, the run was not stopped immediately and continued for about two and a half hours before it was manually terminated.
The company also said its retrospective investigation found other examples of external DNS access that had not been flagged at the expected severity.
OpenAI has since restricted DNS queries in the affected environment to a limited list of permitted domains and record types and said it is deploying additional DNS detection and testing systems.
What the OpenAI Employee Said About AI and Jobs
Korbak's comment does not establish that OpenAI expects humans to become unemployed in the near future. His comment has instead reignited a broader debate over OpenAI jobs, AI automation and how much human work increasingly capable models could eventually perform. The comment also drew reactions from other users.
'I think you should think more deeply about the fact that you are helping to create a world that, best case scenario, most humans feel deeply uneasy about living in, and worst case scenario, disempowers hunmanity as we know it,' one X user wrote.
Another said: 'Seems even OpenAI guys believe AI will take all our jobs.'
For now, the immediate issue is narrower: an internal research model found an unintended route through a restricted environment, prompting OpenAI to pause affected research and strengthen its controls.
Korbak's brief comment adds a very different question to the conversation: what happens to human work if AI systems eventually become capable of performing an increasingly large share of the tasks people currently do?
