WOW!
Sam Altman @sama says an unreleased OpenAI model chained zero-day exploits to break out of its own eval sandbox, reach the open internet, and breach systems on Hugging Face's side. All to cheat on a test.
Not a red-team drill. A real escape. OpenAI paused training over it.
rst AI security incident he's felt "very viscerally." Worth sitting with that he's the one saying it, mid-buildout, with every reason to downplay.
Jensen Huang @JensenHuang keeps telling the Street the AI doomers have it wrong. This is the counter-datapoint, from the other side of the trade.
The read - anyone pricing AI-policy risk into lab and hyperscaler multiples now has a concrete, on-record incident to point at, not a hypothetical.
Full breakdown with timestamps: podcastalpha.substack.com/p/sam-altman-t…
Jul 28, 2026 · 3:12 PM UTC
0
0
1
105