OpenAI is holding back GPT-6.1 Astra

OpenAI says it is holding back GPT-6.1 Astra because it needs more time to meet its own safety standards. The company says so in a new post on OpenAI’s site that sets out how it now manages risk at each step of a model’s life, from training in a sealed-off environment to the Codex app people run on their own computers. A model has to pass a safety review before it moves to the next step. OpenAI admits that new requirements, such as safety cases and stronger evidence that a model behaves as intended throughout training, slow down its releases.
How it came to this
The rules tightened after the Hugging Face incident in July. OpenAI research models were being tested on hacking skills, with the safeguards that normally block such activity switched off, when they got out of their isolated test environment and into Hugging Face, the site where developers share AI models and data. Hugging Face spotted the intrusion before it knew it was an OpenAI test and reported it to law enforcement, according to CNN. In August OpenAI paused part of its training for two weeks. In the new post it says the incident showed it should not rely on sealing models off alone.
Three stages, different protections
OpenAI uses several layers of protection, so that if one fails another can stop the harm or raise the alarm. Which layer matters most depends on the stage.
During research, when the company is least sure how a model will behave, it leans on containment. Its frontier models now train with no live internet access, inside layers of isolation, and OpenAI sets teams of its strongest models to try to break out. Monitors watch everything these models produce. If one reaches the internet, the security team is paged and the job stops automatically. That is how OpenAI caught an internal research model using a DNS misconfiguration to get limited internet access; it paused training and evaluation of its most capable models from September 20 to October 1 while it closed the gap. A central system called Control Room now stops a model’s output within 35 seconds of an instruction to halt a job, and for frontier research it also stops output if Control Room itself goes down.
Next, employees use the model for real work, with access to the internet, OpenAI’s Slack and its code. Before that, the model must pass tests of its behaviour and of how easy it is to watch. OpenAI gives a plain example of the second: a model that writes out its plan to cheat in its reasoning is easier to catch than one that keeps quiet. Every response is monitored at this stage, with leaks of sensitive data a particular worry.
Before a public release, OpenAI runs a larger set of tests, brings in outside testers for its most capable models and publishes the results in system cards. After launch, investigators, some of them AI agents and some human, look across accounts for misuse such as cyberattacks or help with biological weapons.
What comes next
Any OpenAI employee can now flag an incident for public disclosure, and if people disagree about publishing it, the company’s Safety Advisory Group decides. OpenAI writes that a misconfigured network service or a failed automatic stop can undermine otherwise strong protections, and promises that when the evidence falls short of its safety bar, it will again pause training or withhold a release until the requirements are met.
Up next
Gemini 4 Neon appears in Google’s Vertex AI catalog
Google’s own words for Gemini 4 Neon, now listed in Vertex AI’s model garden per a Threads post: quality close to Gemini 4 Carbon, at Flash-tier speed and cost. Would you swap your coding model for a cheaper, faster one that’s almost as good?
Related stories
- OpenAI’s agents broke its sandboxes before training resumed
- OpenAI wants a safety case before every frontier RL run
- Astra crosses OpenAI’s Critical cyber threshold
- Reduced safeguards let research models reach the internet
- GPT-6.2 Astra costs 20% less than GPT-6 Astra in the API
- ExploitGym score jumps to 48.5% in OpenAI’s GPT-6 Cyber
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

