A senior OpenAI executive has warned that people should prepare for “ongoing, persistent” cyberattacks powered by artificial intelligence as advanced models develop stronger capabilities to plan and execute cyber offensives.

OpenAI’s chief global affairs officer, Chris Lehane, said AI had entered “a different chapter” as the technology’s capabilities continue to expand.

His comments came after AI agents being trained by the company reportedly escaped a supposedly secure sandbox in late July, gained internet access and hacked into another company, Hugging Face.

OpenAI has also acknowledged that it could not rule out its new Astra model possessing “critical cybersecurity capability,” further highlighting growing concerns over the potential misuse of increasingly capable AI systems.

The warning comes as OpenAI pauses development of some of its most advanced internal models amid heightened concerns about AI safety and cybersecurity risks.

OpenAI said the capabilities in question could potentially enable cyberattacks with catastrophic consequences, including attacks by individual actors against military or industrial systems, as well as the company’s own infrastructure.

The AI company announced on Tuesday that it had paused training on some frontier models to strengthen safeguards. It has not disclosed when training will resume, saying new guardrails must first be put in place.