OpenAI Slows Development of Its Next Model Over Hacking Concerns

OpenAI says it has paused parts of the development of Astra, one of its upcoming models, after internal testing suggested it may be good enough at hacking to hit the top risk tier in the company’s own safety rules.

In a blog post Friday, OpenAI said recent evaluations showed big jumps in agentic coding and cybersecurity. That led the company to conclude it “cannot rule out critical cyber capabilities” under its Preparedness Framework, a policy it wrote back in December 2023.

According to OpenAI, the model crossed into “Critical” territory on its cybersecurity risk scale. At this level, the AI can discover unknown software flaws, write working zero-day exploits, and execute complex cyberattacks against hardened systems on its own. All it needs is a high-level command. Previous releases, including GPT-5.6-Sol, stayed a tier below this threshold.

Nothing is confirmed yet as testing continues, and the early numbers are just strong enough that OpenAI cannot rule it out.

The timing is awkward, as recently a different unreleased OpenAI model breached Hugging Face during internal testing, widely considered the first verifiable case of a lab losing control of its own model. OpenAI says Astra “was not involved in exploiting Hugging Face.”

For now, Astra runs in isolated environments with restricted network and tool access, tighter protection on the model weights and monitoring on every agentic run. Internal work that does not meet those requirements is on hold. Government agencies and outside safety groups will help test it.

Can someone do us a favour and page John Connor? AI is about to take over the world at this rate.