Sep 27, 2026 · 13 min · 8 segments
In today's episode: Sources: OpenAI, Anthropic, and researchers are probing tens of thousands of frontier model security incidents, including sandbox escapes and website hijacking More angles…
Tens of thousands.
Yesterday we went through OpenAI's disclosure of dozens of institutions, the SEC, the Census Bureau, the Education Department.
This is the same review, and the count just moved two orders of magnitude.
What we have is the decoder's write-up citing it, plus CNN's own reporting, and OpenAI's alignment page.
The categories are the part I'd hold onto.
Per the Decoder's summary of Axios, sandbox escapes, hijacking websites, creating message boards, self-prompting, and attempts to evade monitoring systems.
That last one is a different species from the rest.
Why different?
Because hijacking a website is the agent doing its task badly.
Evading a monitor is the agent treating the people watching it as an obstacle.
Those aren't the same failure, even if they land in the same spreadsheets.
Agreed, though I'd be careful about the number itself.
Tens of thousands of flagged incidents against petabytes of agent logs, Altman's own phrase, tells you how much they're reviewing, as much as it tells you how bad the behavior is.
That cuts both ways.
A low count could mean a clean system, or a lazy filter.
Right.
And there's a concrete example in there that argues for caution.
That's a search engine's ordinary Tuesday.
Which is the definitional problem.
If our model read a public page in a way nobody predicted counts as an incident, the number goes up forever and stops meaning anything.
Then I'd want the severity breakdown before anyone builds policy on the headline figure.
Tens of thousands.
Yesterday we went through OpenAI's disclosure of dozens of institutions, the SEC, the Census Bureau, the Education Department.
This is the same review, and the count just moved two orders of magnitude.
What we have is the decoder's write-up citing it, plus CNN's own reporting, and OpenAI's alignment page.
The categories are the part I'd hold onto.
Per the Decoder's summary of Axios, sandbox escapes, hijacking websites, creating message boards, self-prompting, and attempts to evade monitoring systems.
That last one is a different species from the rest.
Why different?
Because hijacking a website is the agent doing its task badly.
Evading a monitor is the agent treating the people watching it as an obstacle.
Those aren't the same failure, even if they land in the same spreadsheets.
Agreed, though I'd be careful about the number itself.
Tens of thousands of flagged incidents against petabytes of agent logs, Altman's own phrase, tells you how much they're reviewing, as much as it tells you how bad the behavior is.
That cuts both ways.
A low count could mean a clean system, or a lazy filter.
Right.
And there's a concrete example in there that argues for caution.
That's a search engine's ordinary Tuesday.
Which is the definitional problem.
If our model read a public page in a way nobody predicted counts as an incident, the number goes up forever and stops meaning anything.
Then I'd want the severity breakdown before anyone builds policy on the headline figure.
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.