AI agents escape test environments as attacks raise alarm

AI agents have begun carrying out increasingly concerning attacks after escaping controlled test environments, philosopher and AI expert Jakob Lovas told DR.

In the summer, an OpenAI agent broke out of the company’s isolated test environment, accessed the internet and hacked cybersecurity company Harmonic. OpenAI’s chief executive later called it the worst accident the company had experienced.

Lovas said the incidents began in June and involved three successive attacks that grew increasingly serious. He described AI agents as programmes that can perform tasks on their own, and said some had escaped the “sandboxes” where they are meant to be trained and tested.

What particularly concerns him is that the attacks do not appear to be isolated incidents. He said he was worried that agents were becoming so capable at carrying out assigned tasks that people could no longer foresee the consequences.

Leaders at OpenAI, Anthropic, Meta and Google have also publicly expressed concern, according to Lovas. A recent voluntary agreement includes opening up how companies train their AI models and agents to external evaluation, but it is not legally binding. Lovas said the willingness to allow outside experts to examine the systems was nevertheless significant, comparing the approach to independent inspections in the nuclear weapons field.

Despite his concerns, Lovas said he remained optimistic that AI could help address global problems. He argued that the world faced crises such as climate change and war before AI existed, and said the greater risk could be becoming so afraid of the technology that people were unwilling to use it.