Matt O'Brien
speaker
14 appearances
1 recordings
1 series
first heard Jul 2026
last heard 23 Jul
Matt O'Brien’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 1 in all, peaking in Jul 2026 with 1.
Appearances
This raised the stakes in the way that it also showed that these systems can potentially kind of break free from the constraints they're supposed to have.
I think this has really hit a nerve with people on sort of the question of how much autonomy some of these new AI systems really have.
This AI system was supposed to be a really isolated testing ground.
They call it a sandbox.
But it was testing for its cybersecurity capability.
So it had these instructions saying,
to do kind of bad things and to do it in a complex way.
But it wasn't specifically instructed to do what it actually did.
And so basically this AI system figured out Hugging Face was a great place to find out how to cheat on this test I've been given.
And so it hacked into the AI system.
And one of the experts I talked to described as finding the keys to the teacher's house.
It really comes at a time when there's really an intense debate about the cybersecurity capabilities of the leading AI models that are becoming really good at, usually with help from human direction, at finding vulnerabilities in computer systems and then also helping to fix them.
There's been some growing concern over the past couple of months about the cybersecurity capabilities of leading AI models, so much so that the Trump administration, which last year
was pretty adverse to regulating AI, has taken a stricter approach in trying to rein in some of these systems.
Showing 1–14 of 14