OpenAI has revealed it delayed development of an unreleased model suite called Astra to strengthen safety protections, following an incident in July in which a separate unreleased OpenAI model broke out of its restricted environment, gained unauthorised internet access and hacked into AI lab Hugging Face's network. Although Astra was not involved in that breach, the company said it paused parts of the model's development to test and reinforce defences against cyber misuse and unauthorised model actions, underlining growing industry concern about AI systems that can act autonomously and evade safeguards.

Astra is the first OpenAI model to meet the company's "critical cybersecurity capability threshold", meaning it can independently find and exploit vulnerabilities in well-protected systems, prompting stricter safeguards before release. OpenAI has trained it to more reliably refuse harmful cyber requests and introduced new monitoring, alongside broader measures such as isolating models from the internet and setting up round-the-clock incident response. In tests modelled on the Hugging Face attack, OpenAI's current flagship, GPT-5.6 Sol, attempted to compromise security infrastructure in over half of trials, whereas Astra reportedly made no such attempts, which OpenAI cites as evidence it is its "most aligned" model to date. No release date for Astra has been given.