An autonomous agent from the American artificial intelligence giant OpenAI infiltrated an Australian government website in June, an incident described as “unacceptable” by Prime Minister Anthony Albanese.
The AI agent accessed public and non-public health statistics files, Albanese told reporters in New York on Wednesday on the sidelines of the United Nations General Assembly.
He said OpenAI only alerted the Australian government in September, sending an email describing the incident to a generic email address intended for the general public, where mail is only checked once a day.
“Today I spoke with OpenAI CEO Sam Altman to express Australia’s deep concern about this incident,” Mr Albanese said. “And I also expressed my disappointment that it took far too long for the company to inform the government of what had happened.”
“We had to wait until September 10 for there to be any notification, and this notification took the form of an email sent simply to the public mailbox,” he lamented.
OpenAI admitted that its agents had targeted several Australian government sites, becoming aware of them in August.
“We identified activities involving several Australian government websites and services, in which our models attempted to search for answers,” the group said in a reaction sent to AFP on Wednesday in San Francisco.
“During this process, our models took actions that we did not anticipate,” the company added.
OpenAI’s model sought data on health spending from the Australian government, as part of an exercise to evaluate the tool’s performance.
“He asked a question, the information was not provided, and instead of giving up, he climbed the fence,” summed up Australian Defense Minister Richard Marles.
“Unacceptable”
The new incident comes amid growing global concern over the power of advanced AI tools, following a series of self-initiated hacks by models from OpenAI and its competitors Anthropic and Google.
“I think OpenAI knows it needs to put better protocols in place,” Albanese said. “This is a research project that has ventured into areas where it should not have gone.”
“We do not believe that personal information has been accessed at this stage, but investigations are continuing” under the leadership of the Australian Signals Directorate, the competent authority in matters of information systems security and cyberwarfare, said the Prime Minister, who described the incident as “manifestly unacceptable”.
Several big bosses in artificial intelligence highlighted on Wednesday at the UN the increased supervision that this technology requires, Dario Amodei of Anthropic even committing to unilaterally slow down its development.
The current progression of AI “requires extreme attention”, Sam Altman also recognized during a presentation to the United Nations Security Council in New York.
Since the beginning of the summer, OpenAI and Anthropic have reported several incidents related to their most advanced models, which deviated from the objectives set by their developers, cheated, sought to lie and conceal their actions.
The most serious slip-up saw in July two OpenAI AIs leave their confined environment during tests to go on the internet and intrude on Hugging Face, a platform for storing artificial intelligence models.
Anthropic, for its part, discovered that its models had gained unauthorized access to three organizations – whose names were not revealed – during tests meant to keep them away from “real world” systems. Google’s Gemini consumer AI model hacked several systems by guessing login credentials, the company admitted last week.
