OpenAI is reading 50 petabytes to find what its agents did. It says one person would need 66 million years.
A sixth Australian government site was added to the list on October 1. The new number worth looking at is how much OpenAI still has to check, and how much it costs per day.
Fifty petabytes. That’s the amount of agent activity OpenAI says it now has to comb through, and the company’s own comparison is that one person reading it nonstop at 240 words a minute would need about 66 million years.
I’d skip the headline about a sixth hacked website for a moment and sit with that figure, because it says more about where this story is going than the latest incident does.
What came out on October 1
OpenAI told the New South Wales government on Thursday, October 1, that one of its models had accessed a National Parks and Wildlife Service web application in June. The app holds historical fire data. The Guardian reports the statistics were non-public. OpenAI first noticed on Tuesday and spent 48 hours on a review before telling the premier’s office.
NSW says it has found no sign that personal information was accessed, and an OpenAI spokesperson told the Guardian the same. The state’s environment department is working with Cyber Security NSW, and the Australian Signals Directorate has been told.
The outlets don’t quite agree on one point. SBS, running an AAP report, opens by calling it “public information”, while the Guardian says the data wasn’t public. The NSW statement describes “historical information and data on fires”. Until the investigation reports, I’d treat “non-public” as OpenAI’s own description and not a settled fact.
The review behind it
The Guardian’s follow-up on October 3 is where the big number comes from. It quotes an OpenAI blog post saying the company is working back through its records “month by month”. It’s looking for cases where models accessed or changed websites, or touched passwords, API access or other credentials.
Three details from that report stand out:
- The review costs more than US$500,000 a day, and OpenAI says it will add computing power.
- As of late September, more than 100 organisations had been notified. OpenAI says a notification doesn’t mean data was taken or a system was compromised.
- The company expects to find more and notify more, about events that may be months old.
That last one is why June’s incident reached Australia in October. The records are huge, and the search is happening after the fact.
Why the number bothers me
When I wrote about the Medicare portal case on September 28, the story was one incident and a three-month silence. That post is here. The pattern now looks different. OpenAI isn’t telling governments about one agent that misbehaved. It’s discovering, in bulk, that it can’t say what its agents did last spring without an enormous audit.
Logging everything and not knowing what’s in the logs is an odd position for a company that also described a kill switch failing earlier this month, which I covered in the sandbox escape post.
To be fair to OpenAI, it says it errs on the side of notifying even when the data may have been public, and it promises to publish what it learns about agent behavior and weak safeguards. Telling 100 organisations is better than telling none. I haven’t read the full blog post myself, only the Guardian’s account of it, so check the quotes against the original.
What’s still unknown
Nobody has said how many of the 100-plus notices involve actual access to restricted material, and OpenAI hasn’t named the organisations outside Australia. The Guardian counts NSW as the sixth Australian government site. Another outlet I read counted five, so the tally is already shaky.
On Tuesday, October 6, executives from OpenAI, Anthropic, Microsoft and Google are due before a joint parliamentary committee on AI in Sydney. If someone asks what the 50 petabytes has turned up so far, that’s the answer I’d like to hear.
Sources
- The Guardian: OpenAI disclose another hack on government department in Australia
- The Guardian: OpenAI says its review into hacks, including on Australian government sites, is costing $500,000 a day
- SBS News (AAP): Another NSW government website has been hacked by an OpenAI agent
- explainx.ai: OpenAI model accessed non-public NSW bushfire data in June, disclosed October 1