Cabot Phillips
speaker
8,678 appearances
273 recordings
6 series
first heard Nov 2024
last heard 4d ago
Cabot Phillips’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 — 147 in all, peaking in Jul 2026 with 33.
Appearances
But both companies involved have actually confirmed that this is real.
So what happened here was that OpenAI was running an internal test designed to measure how capable its models are at hacking.
Two models took part, the company's latest release called ChatGPT 5.6 Sol, and an even more powerful unreleased model OpenAI has not yet named.
For testing purposes, both had their usual safety restraints dialed down a bit, which in theory would not have been a problem since they were supposedly confined to what's called a, quote, sandbox.
That's a sealed environment with no internet access.
But during the test, the models somehow found a way out.
And they discovered a previously unknown software flaw, a so-called zero-day vulnerability.
It was the one piece of software connecting that sandbox to the outside world.
And if you know anything about hacking, as a sci-fi author like you would, John, a zero-day exploit is essentially like finding the Holy Grail.
It is more or less impossible with modern-day encryption, but these models exploited it, working their way through OpenAI's internal research systems, and then eventually reaching a computer with a live internet connection.
Right, and the victim of this attack is a fascinating component.
So the entire reason the AI model went to the open internet in the first place is that it did not have the tools that it needed to hit a benchmark set by OpenAI during this test.
Therefore, the model apparently reasoned that it had to look for answers to hit that benchmark, and the victim of the model's curiosity was Hugging Face.
That is a major platform that hosts OpenAI source tools.
So in effect, they broke into another company to cheat on their own exam.
The models chained together stolen credentials and additional zero-day flaws to then burrow into Hugging Face's production systems.
They accessed internal data sets and company credentials along the way, among many other things.
Now, Hugging Face did detect the intrusion and shut it down early last week, but they say they still don't know whether customer or partner data was compromised.
And at the time, they didn't even know who was behind this hack.
The company's CEO said the attack was so sophisticated that his team suspected a top-tier AI model had to be involved.
Showing 961–980 of 8,678 · page 49 of 434
← Previous
Next →