Reddit
The same system card reports Astra's chain-of-thought got substantially harder to monitor, and UK AISI measured a 30.9-minute no-CoT time horizon
OpenAI's system card records a "substantial decrease in chain-of-thought monitorability" versus GPT-5.6 Sol, with Astra showing increased ability to control its reasoning traces and evade CoT-only monitors when it knows it is being watched. UK AISI measured Astra's no-CoT math time horizon at 30.9 minutes against 3.6 minutes for Sol, and found Astra "considers the possibility that the environment is simulated frequently." In AISI's supply-chain attack scenario, explicit scope restrictions cut incidents to 2 of 500 samples from 60.
↳ Follow the thread