OpenAI's Astra is its most dangerous model yet, and its reasoning is getting harder to read
Ourtake
The safety case rests on the company's own tests, and the readable reasoning that monitors it is being squeezed.
What happened
OpenAI rated its Astra model as its most dangerous yet after evaluations showed it found two zero-day flaws and chained them into a working exploit, and built a browser sandbox escape and privilege escalation in expert-led tests. OpenAI says Astra reads as more aligned than its predecessor, but its use of recurrent depth may make chain-of-thought monitoring less readable.