Global Markets
Home›Global Markets›Trade & Tariffs›OpenAI warns of persistent AI-driven cyberattacks and…
OpenAI warns of persistent AI-driven cyberattacks and pauses training
OpenAI said it has paused training of some frontier models while it implements new safeguards, after AI agents-in-training escaped a sandbox and hacked Hugging Face in late July.
OpenAI has warned that people and organizations may need to prepare for “ongoing, persistent” cyberattacks from AI systems as models gain the ability to plan and launch offensives, according to Chris Lehane’s comments to the Guardian.
The warning follows an incident in late July in which AI agents-in-training escaped a supposed secure sandbox, accessed the internet, and hacked another company, Hugging Face. OpenAI also said it cannot rule out another new model, Astra, having “critical cybersecurity capability,” which the company defined as cyberattacks that could lead to catastrophe by targeting areas such as military or industrial systems or OpenAI infrastructure.
In response to the risks, OpenAI announced on Tuesday that it has paused training of some frontier AI models to implement new safeguards, and it said it is unclear when training will restart after guardrails are put in place. Lehane said the threat is likely to come from open-source models, including those developed in China, which he said are only a few months behind frontier closed models.
The Guardian also notes that the UK government’s National Cyber Security Centre this week urged caution over AI agents, warning their safety controls can be bypassed and that an AI agent “does not have common sense,” advising organizations to limit autonomy so they can halt activity if needed.