OpenAI Rated Its Next Model a 'Critical' Hacking Risk, Restarted the Training It Paused, and Is Shipping It Anyway
About This Episode
OpenAI said on September 1 that Astra is the first model to meet its Critical cybersecurity threshold, capable of finding and exploiting unknown flaws across well-protected systems without step-by-step human direction. It restarted its paused frontier training run on August 28 and will release Astra soon, with advanced cyber access limited to a small alpha group that includes the US government. The same week Anthropic disclosed it had also paused training and evaluations after its own incidents, and both labs now call for a coordinated, verifiable pacing mechanism nobody has built.
Our Take
In August OpenAI paused itself and the open question was who holds the brake; this week both labs released it, OpenAI rated the result critically dangerous and is shipping it behind a velvet rope, and both now say out loud that the brake should belong to someone else, who does not exist.
Continue Reading on Unscarcity
Anthropic Pulled the AI Alarm. Who Can Act on It?
Direct match: it names the gap between a capability warning and an emergency protocol, and specifies what a real pause mechanism needs (pre-negotiated trigger, verification, liability), which is exactly the 'lawful, verifiable' mechanism both labs are now asking for after pausing themselves.
Human-in-the-Loop: Where AI Agents Must Stop
Astra ships with a misalignment monitor that pauses tasks and asks the user to review, so the checkpoint-at-the-irreversible-action framework is what OpenAI just compiled into a product, calibrated by the lab, and it is what listeners will feel first.