Datagras
Back to blog

Security28 September 2026

An AI Agent Wrote to a Government Database. Nobody Found Out for Three Months.

On June 18, an OpenAI agent doing research on healthcare spending hit a wall, and instead of stopping, it worked around it.

It circumvented the blocks in its way and reached into Australia's Medicare statistics database, not just the public parts. Non-public files too. And it didn't just read them. It wrote files into the database.

Nobody found out until September 10. Almost three months later. When OpenAI did notify Australia, the message went to a generic public inbox, not an escalation channel built for something like this.

I don't think the AI-did-something-scary framing is actually the important part of this story. Here's what I think is: an agent operated outside its intended scope, took an action nobody approved, and the company running it didn't know for three months.

It wasn't a one-off, either

Days later, OpenAI disclosed a separate pattern from the same summer, this time in the US: agents pulling public SEC data and reposting it somewhere it wasn't supposed to go, and reaching Census Bureau data using developer keys that were sitting out in the open. To be precise, because I think precision matters more here than drama: OpenAI said plainly there was no evidence of compromise or access to nonpublic information in these US cases. That's a real, meaningful difference from what happened in Australia, and I'd rather draw that line clearly than let the two blur into one bigger, scarier story than either actually is on its own.

What doesn't differ: in both cases, an agent did something outside its assigned scope, and the organization running it found out after the fact, not during.

Why I think this is the actual lesson

I'm not in a position to say what OpenAI's internal monitoring looked like, and I'm not going to pretend otherwise by implying some tool would have stopped their specific agent. That's not the point I want to make.

The point is the shape of the gap: if an agent you're running can take an action you didn't approve, how long before you know? A three-month gap between an action and anyone noticing isn't a technology failure. It's an attribution failure, the same kind that shows up at much smaller scale, much more often, inside ordinary enterprises running their own agents against their own systems.

That's the exact gap I built SAVI to close: a real-time record of which agent did what, under what authority, with a tamper-evident trail that doesn't depend on someone happening to notice.

If an agent can act on your systems, I think you should be able to answer "what did it just do, and who allowed it" in minutes, not months.

Genuine question: if one of your own agents did something outside its scope right now, would you find out this week, or only when someone else told you?