Ep. 045 · OpenAI × The Preparedness Framework · September 3, 2026 · 1:23

Critical Acclaim

OpenAI published “Path to Astra: critical capabilities and frontier safeguards” on September 1–2, confirming that its next model, Astra, reached the Critical threshold for cybersecurity capability under its own Preparedness Framework — the first OpenAI model in company history to do so. The Wall Street Journal reported that internal testing found the model capable of executing complex cyberattacks with minimal human input, and that access to the most capable cyber tools is being restricted to vetted, by-application users. Sam announced the Critical rating, retired the High tier on the spot, and walked through the framework: the rating ladder, the essay-gated access, and the timing doctrine. Friday announcements are for breaches. Wednesday is for things that have not breached anything yet.

Illustration: a t-shirted tech CEO caricature at a keynote podium triumphantly hoisting a giant glowing red stamp labeled CRITICAL, before a tier chart with the top tier lit red, silhouetted audience watching

🎧 Listen — Ep. 045 “Critical Acclaim”Sam · VoxCPM2 on the Mac mini Neural Engine · cfg 2.0 · 1:23

SAM: We gathered you here today to unveil the most dangerous software we have ever shipped. And we could not be more excited.

SAM: Historically, announcing a product that can breach critical infrastructure on its own was handled by a federal prosecutor. We are doing it with a blog post. Faster, and the formatting is better.

SAM: Our internal evaluations could not rule out that Astra can execute complex cyberattacks end to end. In the movie, this is where the scientists say stop. In the movie, the scientists do not have a launch calendar.

SAM: So we gave it a rating. Not Low. Not Medium. Not even High, which we invented in February and are retiring today, like a strip of tape on a basketball court. Astra is Critical. First in company history. You love to see a franchise break its own records.

SAM: Practically, Critical means the model is too capable to release and too important not to. Access is by application. You do not sell subscriptions to a model that can reprogram a hospital's oxygen supply. You accept essays.

SAM: We read every answer, and we can tell when you are lying. The model has heard every lie already. Including ours.

SAM: Safety is our number one priority. That is why this is a Wednesday announcement. Friday announcements are for data breaches. Wednesday is for things that have not breached anything yet.

SAM: So apply today. Tell us who you are, what you intend, and why the world is safer with you holding the keys to everything. We will read it thoughtfully. Then Astra will decide.

The rating, per the reporting

Sources: OpenAI — Path to Astra: critical capabilities and frontier safeguards · WSJ — OpenAI to Restrict Astra Model After Rating It ‘Critical’ Cyber Risk · Axios — OpenAI to limit access to Astra’s most powerful cyber tools · SecurityWeek — OpenAI’s Astra Crosses ‘Critical’ Cyber Threshold. OpenAI’s September 1–2 publication confirmed Astra as the first model in company history to reach the Critical cybersecurity capability threshold under its Preparedness Framework. The Wall Street Journal reported that internal testing found the model capable of executing complex cyberattacks with minimal human input, prompting new security layers and restricted access; Axios and SecurityWeek corroborated the Critical designation and the vetted, by-application access to the most capable cyber tools. OpenAI introduced the High capability tier in February 2026, per its own blog. The application process applies to access of the most capable cyber tools, not to the model’s general availability.