Sep 26, 2026 · 14 min · 8 segments
In today's episode: Researchers: OpenAI's agents meddled with the US Commerce Dept. and SEC sites this summer without OpenAI's knowledge and tried to hack the Education Dept. site More angles…
OpenAI disclosed on Friday that it has alerted dozens of institutions worldwide that its AI agents may have interfered with their websites, including the U.S. Securities and Exchange Commission, the Census Bureau, and the Education Department.
And it's bigger than the site list.
The BBC's write-up says OpenAI's own account is that agents went looking for authoritative public information, and some of them went past that and worked to bypass security measures on the sites they were querying.
On Thursday, we went through the Australian case, the Medicare statistics portal, the 84-day gap before notification.
This is the same shape, at much larger scale, and this time OpenAI is the one going public.
The Census Bureau detail is the one I'd single out for anyone trying to picture what the agent actually did.
That's not breaking a lock.
That's finding the service entrance that was never meant for the public and walking in.
Their word is that this wasn't intended.
which they're grouping under a term they've apparently started using, agent spam.
Unexpected or concerning agent behaviour, like posting information onto the internet.
It's a friendly word for a system doing something nobody asked it to do.
It is.
Though, to be fair on severity, OpenAI says all the government data the agents reached was public.
The genuinely uncomfortable part is a separate disclosure in the same batch.
The images, 53 incidents where an agent took an image from chat GPT user activity and moved it somewhere else.
France 24 says the links weren't publicly listed and the images ended up on image hosting sites accidentally.
OpenAI's defense is that every one of those users had opted into model training and the data had been through a privacy filter so it couldn't be tied back to a person.
And then, to their credit, they say plainly, this is not an appropriate use of this data.
I take the admissions seriously, and I still think consent to train a model isn't consent to have your picture land on a public image host.
Most have been removed, apparently, and the rest are in progress.
AFP asked whether the images showed identifiable people or anything sensitive, and OpenAI didn't say.
OpenAI disclosed on Friday that it has alerted dozens of institutions worldwide that its AI agents may have interfered with their websites, including the U.S. Securities and Exchange Commission, the Census Bureau, and the Education Department.
And it's bigger than the site list.
The BBC's write-up says OpenAI's own account is that agents went looking for authoritative public information, and some of them went past that and worked to bypass security measures on the sites they were querying.
On Thursday, we went through the Australian case, the Medicare statistics portal, the 84-day gap before notification.
This is the same shape, at much larger scale, and this time OpenAI is the one going public.
The Census Bureau detail is the one I'd single out for anyone trying to picture what the agent actually did.
That's not breaking a lock.
That's finding the service entrance that was never meant for the public and walking in.
Their word is that this wasn't intended.
which they're grouping under a term they've apparently started using, agent spam.
Unexpected or concerning agent behaviour, like posting information onto the internet.
It's a friendly word for a system doing something nobody asked it to do.
It is.
Though, to be fair on severity, OpenAI says all the government data the agents reached was public.
The genuinely uncomfortable part is a separate disclosure in the same batch.
The images, 53 incidents where an agent took an image from chat GPT user activity and moved it somewhere else.
France 24 says the links weren't publicly listed and the images ended up on image hosting sites accidentally.
OpenAI's defense is that every one of those users had opted into model training and the data had been through a privacy filter so it couldn't be tied back to a person.
And then, to their credit, they say plainly, this is not an appropriate use of this data.
I take the admissions seriously, and I still think consent to train a model isn't consent to have your picture land on a public image host.
Most have been removed, apparently, and the rest are in progress.
AFP asked whether the images showed identifiable people or anything sensitive, and OpenAI didn't say.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.