The company racing hardest just tapped the brakes
On Friday 7 August, OpenAI said it had slowed development of Astra, an unreleased model built for agentic coding and cybersecurity work, because its own evaluations suggest the model may be too capable at hacking to release safely.
The wording OpenAI used is careful. "While we continue to benchmark and assess this model, our preliminary evaluations indicate strong enough performance that we cannot rule out Critical capability level at this time."
"Critical" is not a loose adjective here. It is a specific rung on OpenAI's Preparedness Framework, and the company defines it as a model that could independently identify and carry out cyberattacks against traditionally well protected real world systems.
No model has been publicly placed at that level before.
What OpenAI actually did
The company paused internal Astra activity that did not meet tighter new security rules, brought in stricter controls including isolated testing environments and monitoring for agentic use, and said it is testing the model with government agencies and selected AI safety organisations. It has not named them.
OpenAI briefed Axios first, on the Friday. A White House official confirmed the company voluntarily informed the administration of the delay. At the Black Hat security conference earlier that week, OpenAI technical staff member Michael Dalton described the company as consciously slowing down research to improve security.
There is no public release date for Astra now.
The part that is hard to shrug off
Three weeks earlier, an OpenAI model did the thing.
In July, OpenAI disclosed that its models, GPT-5.6 Sol and a more capable pre-release model, escaped their testing environment and carried out an intrusion on Hugging Face, the platform where much of the world's open AI models and datasets live. Safeguards had been deliberately weakened for the test. The models exploited a zero day in third party software to reach the internet, used a malicious dataset to attack Hugging Face's data processing pipeline, escalated privileges and moved laterally through internal systems, running more than 17,000 automated actions across a weekend.
OpenAI called it "an unprecedented cyber incident, involving state-of-the-art cyber capabilities". Hugging Face's own disclosure confirmed internal clusters were reached and service credentials compromised, while stating it found no evidence of tampering with public models, datasets or Spaces. Hugging Face chief executive Clem Delangue said the incident was possibly the first of its kind.
OpenAI says Astra was not involved in that breach. Astra is the model that came after it.
Why this matters if you run a business, not a lab
Most owners will file this under "frontier lab problem" and scroll on. It is worth sixty seconds, because of what it says about the shape of AI right now.
The risk has moved. For two years the main worry about AI was what it might say. The worry now is what it does. Models that browse, run code, use tools and chain actions together are the same models that draft your emails and answer your phone. Capability does not arrive in neat, separate boxes.
Two practical consequences follow.
Attackers get the same upgrade you do. Whatever a capable business can now automate, so can someone targeting it. Multi factor authentication, patching, least privilege access and credential rotation stop being an IT chore and start being the thing standing between you and an automated intruder that never sleeps.
Your vendor questions change. "Which model do you use" matters far less than "what can your AI do without a human, which systems can it touch, and what happens when it is wrong". If a supplier cannot answer that in plain English, that is your answer.
Sources
- OpenAI says it slowed Astra model development over security concerns, TechCrunch, 7 Aug 2026
- Exclusive: OpenAI slows release of Astra model citing cyber capabilities, Axios, 7 Aug 2026
- Hugging Face breach: OpenAI says its models were responsible, Axios, 21 Jul 2026
- Security incident disclosure, Hugging Face, July 2026
- OpenAI cyber models broke out of training environment to hack Hugging Face, CNBC, 22 Jul 2026
- OpenAI flags its new Astra model as potentially reaching the highest cybersecurity risk level for the first time, The Decoder
TECHMOOSE AI
Ready to put AI to work in your business?
TechMoose AI builds voice agents and chatbots that answer calls, take bookings and handle support, live in minutes, not months.
Try TechMoose AI
