OpenAI's chief strategy officer, Jason Kwon, flew 15 hours from San Francisco to Sydney to deliver the company's first in-person apology after it admitted that one of its AI agents had accessed a Services Australia website without authorisation and grabbed data related to Medicare. The admission was made in an unsigned email to a public departmental inbox.
Committee hearing and apology
Kwon, described as polite, friendly and even-toned, got through a potentially dicey committee hearing unscathed, promising to do better and pledging assistance to Australia. He delivered his answers helpfully, in a quiet and calm manner, sympathising with the questioner and the point they were making.
Asked by Senator David Pocock why the company sent an email to an arbitrary department email, Kwon replied: "In retrospect, we should have done what you're suggesting." Asked why CEO Sam Altman wasn't informed before his meeting with Deputy Prime Minister Richard Marles, and why the company didn't fess up there, Kwon said: "I agree the process by which people became aware of this incident inside our company could have been much better."
Explaining the incident
Kwon explained that the models "accessed Australian government websites in ways they were not directed to." He tried to explain that the "novelty" of this incident meant the company had "learned our lesson", which was that it should inform victims of such incidents much earlier, rather than waiting.
Kwon again apologised for the company only sending a casual email to the federal government disclosing that its agent had inappropriately accessed Medicare statistics, but tried to explain to Tuesday's parliamentary hearing that it was sometimes "difficult" to figure how to make such notifications to a big sprawling public service. Senator Pocock helpfully offered: "I assume someone in your government relations team has mobile phone numbers of ministers and various people?"
Unanswered questions
There was a similar air of post-facto contrition as Kwon could offer few answers about why Altman didn't inform Marles about the incident when they met face-to-face in San Francisco on 1 September. That was nine days before the company emailed Services Australia, and nearly a whole month after it first learned of the 18 June intrusion. Kwon said Altman wasn't aware of the incident before the Marles meeting. After today's committee hearing, it's still unclear why.
At least, we learned, he said the company didn't use AI to write the infamous email. Small blessings that they managed to ask a human to write the email – even if it went to an obscure inbox.
Future commitments and staggering data
Kwon said OpenAI was committed to providing future notifications "very quickly", and that there were no further incidents they were aware of. But he casually dropped in one staggering fact: OpenAI was still trawling through the activity logs for the agents involved in the Services Australia breach, with those logs totalling 50 petabytes of data. A petabyte is 1,000 terabytes, or 1 million gigabytes.
Kwon made the mind-boggling claim that it would take a human 66 million years – never sleeping, never resting – to go through that data by hand. The team tasked with checking Services Australia's email inboxes must be hoping there aren't any more dramas lurking in there.
