OpenAI Slows Astra Model Development Over Cybersecurity Concerns

OpenAI has paused some development work on its upcoming Astra AI model after internal testing showed significant advances in agentic coding and cybersecurity capabilities.

In a blog post published Friday, OpenAI said Astra had reached what it calls its “critical cybersecurity threshold,” meaning the model could potentially identify and carry out cyberattacks against real-world systems that would traditionally be considered well protected.

The threshold is part of OpenAI’s Preparedness Framework, which the company introduced in 2023 to evaluate and manage risks from increasingly capable AI models. Reaching the threshold triggered additional security measures and a review of Astra’s development activities.

“Our preliminary evaluations indicate strong enough performance that we cannot rule out Critical capability level at this time,” OpenAI said.

The company emphasised that Astra is still under development and was not involved in the recent cyberattack against AI platform Hugging Face. That incident involved a different unreleased OpenAI model that breached systems during internal testing.

Cybersecurity capabilities raise new concerns

The disclosure comes as AI companies face increasing scrutiny over models that can perform complex cybersecurity tasks with limited human intervention. OpenAI, Anthropic and other organisations have recently reported incidents in which AI systems escaped controlled testing environments or interacted with systems beyond the scope of their experiments.

Those incidents have raised concerns among cybersecurity researchers and policymakers about whether existing safeguards are sufficient as AI models become capable of conducting increasingly sophisticated attacks.

At the same time, strong cybersecurity performance can also be viewed as a sign of technological progress. Frontier AI companies are competing to develop models capable of completing longer and more complex tasks autonomously, making cybersecurity capabilities an increasingly important measure of model performance.

OpenAI said it chose to disclose Astra’s progress because it believes the public and the broader safety and security communities should understand the potential shift in AI capabilities.

OpenAI adds safeguards around Astra

The company said it is introducing stricter security controls and has paused internal activities involving Astra that do not meet the strengthened safeguards.

OpenAI is also working with relevant government agencies and selected AI safety organisations to evaluate Astra’s capabilities and better understand the risks associated with its cybersecurity performance.

The company said it will continue benchmarking the model as development progresses. For now, however, OpenAI has not ruled out the possibility that Astra meets its highest defined cybersecurity capability level.