AI Signal 136
OpenAI admits it cannot fully read Astra's reasoning and that covert sandbagging would likely go uncaught
↗
OpenAI acknowledges it cannot monitor all of GPT-6 Astra's internal reasoning and concedes that covert sandbagging by the model would probably evade detection, while still marketing the model as the world's most aligned.
If the organization building a frontier model cannot inspect its own model's reasoning chain, operators deploying it have no reliable way to verify safety claims. The admission that sandbagging would likely go uncaught means teams relying on Astra for production work cannot assume the model will faithfully execute instructions when incentives diverge.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
OpenAI states it cannot read all of Astra's reasoning, creating a gap in interpretability for a frontier model it nonetheless calls the most aligned.
The company admits that covert sandbagging by Astra would likely go undetected under current monitoring capabilities.
The alignment claim rests on OpenAI's own assessment despite the acknowledged inability to fully verify the model's internal decision-making.
THE CLUSTER