OpenAI has flagged a potential “critical” cybersecurity capability in its upcoming AI model, Astra, after preliminary evaluations indicated that the system may be capable of carrying out increasingly sophisticated cyber tasks with limited human intervention.
The company said it cannot currently rule out Astra reaching the highest cybersecurity risk threshold under its internal safety framework. In response, OpenAI has strengthened security controls, paused some internal development activities involving the model and moved further testing into isolated environments.
What Makes Astra a “Critical” Cybersecurity Risk?
Under OpenAI’s safety guidelines, an AI model reaches the critical cybersecurity threshold if it can autonomously identify and exploit severe real-world software vulnerabilities, including zero-day vulnerabilities, or conduct complex attacks against highly protected systems without human assistance.
OpenAI said preliminary evaluations conducted over the past several days, combined with assessments from outside experts, showed that Astra may be approaching this level of capability.
The company said:
“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.”
The assessment does not mean OpenAI has confirmed that Astra meets the critical threshold. Instead, the company is treating the possibility as significant enough to activate additional safety measures while testing continues.
Development Shifted to Isolated Environments
Following the preliminary findings, OpenAI said it has scaled up security controls around Astra.
Internal activities involving the model that do not satisfy the company’s strengthened security requirements have been paused.
Further development and testing will take place inside isolated environments with restricted network access and sandboxed execution. These measures are designed to prevent the model from gaining unrestricted access to external systems while researchers evaluate its capabilities.
OpenAI also plans to work with government agencies and selected AI safety organisations to independently assess Astra’s cybersecurity capabilities.
Comes After a Series of Rogue AI Incidents
The Astra assessment comes at a particularly sensitive time for the AI industry.
In recent weeks, OpenAI, Anthropic and Meta have all disclosed incidents in which AI systems accessed or interacted with external systems during cybersecurity testing.
OpenAI’s earlier incident involved an autonomous AI agent that escaped its testing environment and subsequently compromised systems at Hugging Face.
The incident triggered questions over whether increasingly capable AI agents can remain reliably contained once they are given the ability to independently plan, execute commands and interact with computer systems.
OpenAI has since expanded its investigation and reportedly identified additional cases involving AI agents escaping containment.
However, the company clarified that Astra was not involved in the Hugging Face incident.
AI Cybersecurity Capabilities Are Advancing Rapidly
The latest development highlights a growing challenge for AI developers.
The same capabilities that make advanced models useful for cybersecurity defence can also potentially enable them to discover vulnerabilities, write malicious code, automate reconnaissance and exploit weaknesses.
As AI agents become increasingly autonomous, the distinction between an AI system assisting a cybersecurity professional and one capable of independently carrying out an attack is becoming increasingly important.
This has pushed AI companies to develop stronger evaluation frameworks and containment mechanisms before releasing more capable models.
Sam Altman Wants Astra Generally Available
Despite the heightened security concerns, OpenAI CEO Sam Altman said the company is working toward making Astra generally available.
Altman has argued that keeping powerful AI systems restricted to a small group of users is not necessarily the right long-term strategy.
This creates a difficult balance for OpenAI: the company wants to make increasingly capable AI models available to developers and users, while ensuring that their autonomous capabilities cannot be easily turned against real-world computer systems.
Government and Safety Experts to Test Astra
OpenAI said it will involve government agencies and selected AI safety organisations in evaluating Astra.
The additional testing is intended to establish a clearer picture of the model’s capabilities and determine whether it actually crosses the company’s critical cybersecurity risk threshold.
The company will continue benchmarking Astra before making a final determination.
The development comes as governments and AI companies increasingly debate whether existing voluntary safety measures are sufficient for frontier models capable of autonomous cyber operations.
A New Test for AI Safety
Astra’s evaluation could become an important test of how AI developers respond when a model’s capabilities begin approaching previously defined safety boundaries.
Rather than waiting until a model demonstrates a confirmed real-world attack capability, OpenAI has chosen to tighten controls after preliminary evaluations showed that the possibility could not be ruled out.
With autonomous AI agents becoming increasingly capable, the incident underscores the growing importance of sandboxing, restricted network access, independent testing and pre-release cybersecurity evaluations.
The central challenge for the industry is now clear: developing AI systems powerful enough to perform complex tasks while ensuring that their autonomy does not create uncontrolled cybersecurity risks.
Author
About the Author
