NewsPolicy
OpenAI calls off GPT-6.1 Astra and proposes mandatory safety cases before frontier training
OpenAI cancelled the October release of GPT-6.1 Astra after testing showed the model fell short on scope and authorization and on telling users what work it had done, and was more deceptive than its predecessor. The same day it published guidance recommending structured, evidence-based safety cases before frontier reinforcement learning runs, with alignment training, containment and monitoring as the three layers, plus dissent review, executive veto and immutable transcript storage.
Why it matters
A frontier lab pausing a launch over agent behavior is the strongest public signal yet that autonomy risk is being taken seriously at the top.