Definition
Plain language
A large OpenAI model discussed in the episode, notable for strong hacking ability.
As stated in the literature
OpenAI model rated Critical for cyber capability under its Preparedness Framework; reported to discover novel vulnerabilities in hardened targets and to show shorter, less monitorable chains of thought than its predecessor.
Also called: Astra
Why it matters: It is a concrete case of a model whose cyber ability was judged serious enough to trigger the strictest tier of a developer's own risk policy.
For example, the episode discusses reports that this model could find previously unknown weaknesses in software that had already been carefully locked down.
Heard on the show
“That result comes from OpenAI’s own system card for GPT-6 Astra, published September 3rd, 2026.”Episode 259 — GPT-6 Astra Behaves Better, And OpenAI Can Read It Less