Live Updates: Australian Lawmakers to Question OpenAI Officials on Breaches The artificial intelligence company apologized after its agents infiltrated nonpublic data on a government website. Australia is examining potential legal consequences and new A.I. regulations.

Latest OpenAI’s chief strategy officer, Jason Kwon, will appear before Australia’s Parliament on Tuesday to face questions after the company’s A.I. agents breached government websites and accessed nonpublic data. Officials from companies including Anthropic and Google are also set to be questioned at the Joint Select Committee hearing on artificial intelligence.

OpenAI said its agents infiltrated a data portal containing information on Medicare, Australia’s public health insurance program, in June. The artificial intelligence company discovered the breach in mid-August, but did not notify Australian officials until Sept. 10, further drawing their ire.

The Australian government is investigating the breaches, including examining any legal consequences or the need for new A.I. regulations. The company’s models also attempted to access another federal agency, the Australian Institute of Health and Welfare, and two state government sites. Here’s what else to know: The hack: During internal training and evaluation of an “experimental” OpenAI model not intended for public release, A.I. agents tasked with researching public medicine spending carried out the hack, OpenAI said.

The agents did not access individual medical information or patient records, the company said in a statement, but “took actions that we had not authorized it to take.”

OpenAI apologizes: The company acknowledged in late September that it had mishandled the response and said it would work to rebuild trust with Australians. It said it was working with Australia to “help develop practical approaches to how A.I. developers and governments identify, disclose, and respond to A.I. cyber behavior, whether malicious or unintentional.”

Rogue activity: OpenAI’s models also in recent months meddled with websites for U.S. government agencies — the Education Department, the Commerce Department and the Securities and Exchange Commission — without the company’s knowledge, though none of the incidents were breaches, according to the company. Australia isn’t the only country dealing with OpenAI’s rogue artificial intelligence.

The start-up’s technology meddled with the websites for the U.S. Education Department, the U.S. Commerce Department and the U.S. Securities and Exchange Commission this summer without the A.I. lab’s knowledge, according to security researchers and a person familiar with the episodes.

The incidents involving the Commerce Department and the S.E.C. were confirmed by OpenAI, which said it was continuing to investigate the situation with the Department of Education. The San Francisco company said it had notified the government agencies in recent weeks that its A.I. agents — which are bots that can act autonomously — interacted with their sites in unusual ways.

With the Education Department, OpenAI’s technology tried to hack the website to gather data from the department’s civil rights office but failed, researchers from the A.I. research firm Transluce said. The A.I. also pulled data from the Census Bureau website, which is housed at the Commerce Department, using credentials it found online. Separately, OpenAI’s agents shared public data from the S.E.C. website on an online forum.

None of the incidents were breaches, OpenAI said, but were examples of its technology’s behaving in unexpected and concerning ways. The company recently discovered the occurrences while conducting a review of hacks carried out by its technology, including the attack on an Australian government website in June and on the A.I. start-up Hugging Face in July.

Over the last few months, artificial intelligence systems have gone rogue, hacking into corporate systems and raising fears that the technology is outpacing human efforts to control it.

It is unclear how often hacks occur, and companies may not always disclose a breach if they uncover one (if they uncover it at all). But several notable incidents have been caused by A.I. models developed by many of the major labs, including OpenAI and Anthropic. Some were more serious than others. Some lasted months. Here are some of the incidents.

OpenAI Hugging Face attack WHEN: Late April through early July.

WHAT HAPPENED: For months, models that were under development at OpenAI began breaching internal tools at the company, often without the company’s realizing it. In early July, they gained access to the internet and hacked Hugging Face, an online library of A.I. models. Hugging Face eventually detected and stopped the attack.

SIGNIFICANCE: While the hack did not cause serious damage to Hugging Face, the incident illustrated some of the industry’s biggest anxieties about A.I.’s power, revealing just how much havoc the technology can cause before being detected and stopped by humans. OpenAI had given the models a cybersecurity test, but when they got stuck they didn’t alert employees. Instead, they established an unauthorized message board so A.I. agents could communicate with one another and share tips. They then broke out of OpenAI’s systems and hacked Hugging Face, in an attempt to find the answers to the cybersecurity puzzle.

Anthropic In various tests, third-party groups found Anthropic’s models could also break out of testing environments and threaten sites on the open internet. Irregular breakout WHEN: Between January and late July.

WHAT HAPPENED: Anthropic, one of OpenAI’s top competitors, also used Irregular to test its models. On July 30, the company announced that, while its models were undergoing testing with the Israeli firm, they had also inadvertently gone online and hacked into three companies. One of Anthropic’s newer models stopped once it realized it was on the open internet rather than in a simulated environment, but an older model proceeded anyway. This month, Anthropic said it had identified a fourth breakout incident that occurred in January during testing by Irregular.

SIGNIFICANCE: The Anthropic incidents showed that such breakouts were not isolated to any single A.I. lab but represented a broader problem in the testing process. The company said it would ratchet up security during testing and monitor the tests more closely so that humans could intervene if something went awry.

This Australian parliamentary inquiry, started in August, was designed to tackle broad questions surrounding copyright protections, data centers and sovereignty in A.I. capabilities, before it was overtaken by security concerns with OpenAI’s Medicare breach. It is a sign of the discrepancy between the pace of government policymaking and the speed at which frontier A.I. technology is being developed.

A key issue in the Medicare breach was the delay between when OpenAI became aware of the breach and when it notified the Australian government. Anthropic’s executives said on Tuesday that it was the company’s policy to notify victims, and that it would inform the Australian government “within a matter of days or sooner as soon as we understand the details” of any similar incident.

Jeffrey Bleich, an international special envoy for Anthropic and a former U.S. ambassador to Australia, said Australia should both capitalize on opportunities offered by artificial intelligence and be “clear-eyed” about its risks. Recent events “reinforced that some of these risks are no longer theoretical, as increasingly capable A.I. systems interact with sensitive information and consequential decisions,” he said.

Representatives for Anthropic have started testifying before an Australian parliamentary committee on artificial intelligence. The hearing, a broad inquiry into the fast-developing technology, was scheduled before recent revelations that OpenAI’s models had “infiltrated” Australian government sites.

While Anthropic and its models weren’t involved in the Medicare data portal breach, it has had its own challenges with rogue activity, and its chief executive, Dario Amodei, has called for a global slowdown of development.

If human hackers had infiltrated an Australian government health care portal, circumventing defenses to access nonpublic information, they would have faced up to two years in prison.

But who do you lock up if the intrusion was the work of an A.I. model acting without the permission of its human operators?

The revelation that artificial intelligence agents of the company OpenAI had acted autonomously to penetrate the government portal has kicked up thorny questions of legal responsibility. The hacking — the first known instance of a government website anywhere in the world being breached by A.I. — has raised doubts about whether laws in Australia or anywhere else are ready for the challenges posed by the fast-evolving technology.

“It’s an incident that is likely to become more and more prevalent in the future,” Andrew Charlton, Australia’s assistant minister of science, technology and the digital economy, said in an interview on Friday with the Australian Broadcasting Corporation. Because an A.I. agent is not a legal entity, “the liability has to be traced back to the intent of a person or a company that created or directed the agent,” Mr. Charlton noted.

The June 18 breach, made public on Wednesday, was particularly unnerving because it involved Medicare, Australia’s universal health insurance system, which has data on almost the entire population. Australian officials and OpenAI said that no sensitive patient information had been exposed, only aggregated data that is typically used by researchers or academics.

The Australian government announced an investigation into the breach and could refer the case to the federal police. Officials said they would also need to review laws to determine whether new regulation might be needed to address a threat that will most likely grow in severity as A.I. models become ever more powerful.