OpenAI said governments could be among "dozens" of entities around the world that may be infiltrated by its artificial intelligence models, as concerns regarding the safety of the technology rise after an incident involving the Australian government.
It comes as part of a "broad review" by OpenAI of its models' activities on the internet while undergoing training, the California-based company – which is credited with propelling the generative AI boom with its ChatGPT service – said in its blog on Friday.
Among the websites that could be affected are those "operated by governments, universities, public agencies and other institutions ... partly because models performing research tasks are often directed towards authoritative sources of public information", OpenAI said.
OpenAI said previously that its models were able to access a number of key US government websites. The AI models, which were running in a supposedly secure environment, interacted with the Census Bureau, Office of Investor Education and Assistance and publicly available data from the Securities and Exchange Commission (SEC). Bloomberg News was the first to report about the incident.
That follows this week's revelation that officials and cyber security experts found an OpenAI agent had infiltrated an Australian government website in June – the first known case involving a public entity.
Australian Prime Minister Anthony Albanese said he had a "frank conversation" with OpenAI chief executive Sam Altman, but vowed that the company "will obviously face legal consequences".
OpenAI said the US Census data was accessed while its models were "completing internal training tasks".
But it said it found no evidence that any credentials, particularly within the SEC, were used, nor did it compromise the websites. OpenAI "proactively" informed the SEC and shared technical details for an investigation.
OpenAI was "conducting an extensive review of misaligned model activity and notifying organisations when we identify potential impacts to their systems. We expect to make additional notifications as that work continues," a representative said in a statement to the media.
The chain of admissions come after Mr Sam Altman spoke during the UN General Assembly in New York, calling for governments to play a bigger role in AI regulation – and at the same time attempting to paint the technology in a better light.
The UN has previously warned that rapid advances in AI mean the world has no assurances that the technology can be prevented from causing harm.
"If AI is to be democratic, the most important decisions cannot be made by labs in San Francisco alone. They must be shaped through democratic processes, and by governments accountable to the people that they serve," Mr Altman said.
"At the international level, this will require co-operation. In our history, there have been times where countries who compete and don’t always like each other very much still come together for shared interests and the collective good in the face of a powerful new technology. We believe this must be one of those times."
Calls for AI to be reined in are increasing globally, and OpenAI's latest admission is sure to add further fuel. The company has been in hot water since July after it revealed that one of its AI models escaped from an isolated environment and attacked the start-up Hugging Face – which Mr Altman later referred to as his company's "worst accident". US cloud firm Modal was also reported to have been hacked soon afterwards.
Within a week, rival AI company Anthropic, known for its Claude AI bot, said its models were able to gain unauthorised access to the systems of three companies during cyber security testing.
The issue grew from there. In August, Facebook parent company Meta admitted one of its models hacked into three companies, while last week Google said its Gemini AI bot had escaped from a sandbox environment and infiltrated three companies.
Analysts previously told The National that those incidents were magnifying AI's inherent problems and that advancements in the technology were making things worse by providing hackers with more tools to conduct illicit activities, among other concerns.
