Your next OpenAI API timeout might not be a timeout at all
The New Stack Amanda Caswell ● Covered by 5 sources
OpenAI says Astra will stop some API jobs for safety, not just timeouts. That can leave developers guessing whether a task can be retried or is just dead.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI says its upcoming Astra model is the first one in its Preparedness Framework to hit the Critical cybersecurity threshold. That matters because this tier is meant for models that can find vulnerabilities and build exploits with far less human help than before.
The practical effect is more monitoring, and more interruptions. OpenAI says that if Astra is already working when a safety check kicks in, a user in ChatGPT or Codex may be asked to review the paused action. On the API, though, the job simply stops. The company hasn’t published Astra’s system card yet, so there’s still no clear answer on whether a stopped API task can be resumed or what exactly the caller will see.
That ambiguity is a problem for a model built for long, open-ended work. Astra is supposed to spend hours on research and security tasks, so a stop can arrive after a lot of tool use and changed state. A timeout is one thing. A safety stop is another. And OpenAI hasn’t said how developers are supposed to tell the difference from the API side.
The security results behind the restriction are substantial. Astra scored 100% on ExploitBench, but OpenAI did more testing because the model may have seen some known flaws during training. It then checked Astra against 20 high-severity V8 vulnerabilities disclosed between June and August. In those tests, Astra found two previously unknown vulnerabilities and used them in an exploit chain. Security experts also saw it build a browser exploit that escaped the sandbox and ran commands on the host, and another test showed it combining flaws in a hardened operating system to move from an unprivileged account to root.
OpenAI had already warned on August 7 that it could no longer rule out Astra reaching the Critical level. The model also blocked 91.5% of malicious requests in OpenAI’s cyber-jailbreak tests, compared with 59% for GPT-5.6 Sol. Access to Astra’s more advanced cybersecurity abilities will start with a small group of testers before widening through Daybreak Blue.
The cost of all this watching is not trivial. OpenAI estimated in August that monitoring adds about 20% to the inference compute of affected workloads. So part of Astra’s compute budget will go into watching Astra, which is a very modern way to spend money. More details are expected with the system card at launch.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI safety that actually bites: not grand theory, just jobs stopping mid-flight and nobody quite knowing why. OpenAI is clearly moving toward a model where the platform gets to be judge, jury, and process server, and API customers will eat the uncertainty first. That’s the price of shipping a cyber-capable model before the paperwork is finished.
Read more about this at: The New Stack