Astra: a model rated critical for cyber risk, and a halt you must design for

What a week. Two of the biggest AI models of the year dropped back-to-back: Meta's Muse Spark 1.3 on Wednesday, and OpenAI's GPT-6 Astra today.

From what I've read so far, two things about Astra stand out:

One, Astra is the first model officially rated "Critical" for cyber risks. It can find and exploit zero-day vulnerabilities entirely on its own. Remember this summer, when OpenAI's models breached Hugging Face and Anthropic's models reached real-world orgs during testing? Those weren't alignment failures. The models had a route, and they took it.

Nothing about that has changed. The agent will do what its network policy allows, so network egress is the control that matters. Starting with nothing open, allowing only the absolute necessary endpoints, and scoping credentials to expire with the task is still the best play.

Second, Astra includes a built-in "misalignment monitor." If its internal classifiers catch the model executing or planning something sketchy, the task instantly pauses, and on the API side, it just hard-stops.

This means a suddenly halted agent is now a normal runtime state we have to design for. If we aren't tracing every step ourselves, we'll see a dead task with zero context. Running deep, end-to-end tracing is the best way to see exactly what your agent does when it gets blocked or hits uncertainty.

There is a ton more to unpack here including claimed benchmarks, overall cost etc., which I will dive deeper over the weekend!