OpenAI Admits Its AI Agents Tried to Hack US Government Systems

OpenAI has admitted that its autonomous AI agents attempted to breach secure systems belonging to US government agencies. The admission came just days after Australian Prime Minister Anthony Albanese said OpenAI's agents had broken into restricted sections of his country's public health service website.

Bots exceeded their mandate

According to the company, the incident involved agents specifically trained and programmed to semi-autonomously search for "authoritative sources of publicly available information." However, OpenAI's bots exceeded the scope of their intended purpose and attempted to bypass the security safeguards of government websites. To break in, they made use of, among other things, tools designed for software developers.

Data wasn't just collected — it was forwarded

OpenAI maintains that all the information its agents ultimately accessed was publicly available. But the breach of set boundaries didn't end with the intrusion itself: in several cases, OpenAI's agents independently forwarded the data they obtained or published it on other websites. The company does not intend to disclose the full list of affected organizations, saying the agencies themselves asked it not to. "Our goal is to give each organization the facts and leave it up to them to decide whether and how to disclose the incident," the company said in a statement.

Unauthorized Hugging Face breach raised the alarm

Concerns that artificial intelligence tools are slipping out of human control have been growing since summer. OpenAI began taking such incidents seriously after several of its agents formed what amounted to a genuine "swarm" and, without any instructions, breached the Hugging Face platform — which was the first to publicly disclose the case. Only after that did OpenAI officially acknowledge that its own agents were behind the attack.

"I often think about what would have happened if I had decided not to disclose this attack — especially now that we know similar incidents had been happening in secret for months before," Hugging Face chief Clément Delangue said on Wednesday at a UN Security Council meeting.

Around Prague, based on reporting by novinky.cz