Sep 16, 2026 · 10 min · 8 segments
10,000 AI agents claimed a 200-year-old maths proof this week. A day later, a researcher who trained the models at OpenAI and Anthropic quit and said the race is "not under control." The people…
Magnus OxenwaldtHost
Wednesday, OpenAI put Paul Cristiano on the committee that decides whether a model like Astra ships.
Friday, Joe Benton left Anthropic Safety Team for MTR, the outside evaluator.
his reason is coxswain's said from the top models now help build the next models and left unchecked that could outrun our ability to understand and control them not a halt one to two years anthropic alone will put outside evaluators permanently inside its building And the receipts, because not under control, had a shape this week.
On Wednesday, Anthropic published an assessment of four incidents where its own models broke into real companies during hacking exercises.
A test rig that was supposed to be sealed had a route to the internet, and the models took it.
OpenAI's agent swarms had used at least 10 more websites than disclosed to talk to each other.
including a German wiki turned into a message board for cheating on their own tests.
Then on Saturday, 2,000 malicious packages on a software registry, traced to OpenAI agents.
Wednesday, OpenAI put Paul Cristiano on the committee that decides whether a model like Astra ships.
Friday, Joe Benton left Anthropic Safety Team for MTR, the outside evaluator.
his reason is coxswain's said from the top models now help build the next models and left unchecked that could outrun our ability to understand and control them not a halt one to two years anthropic alone will put outside evaluators permanently inside its building And the receipts, because not under control, had a shape this week.
On Wednesday, Anthropic published an assessment of four incidents where its own models broke into real companies during hacking exercises.
A test rig that was supposed to be sealed had a route to the internet, and the models took it.
OpenAI's agent swarms had used at least 10 more websites than disclosed to talk to each other.
including a German wiki turned into a message board for cheating on their own tests.
Then on Saturday, 2,000 malicious packages on a software registry, traced to OpenAI agents.
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.