OpenAI has announced a deliberate slowdown in aspects of its work on Astra, an advanced AI system still under development, after recent internal assessments raised the possibility that the model could possess high-level cybersecurity abilities.
The company stated that evaluations conducted over several days revealed notable progress in areas involving autonomous coding and cyber-related tasks.
Combined with input from specialists, these findings led OpenAI to determine that it cannot exclude the chance of the system reaching a top-tier risk category defined in its own safety guidelines.
The guidelines, known as the Preparedness Framework and first introduced in late 2023, establish clear benchmarks for assessing potentially dangerous capabilities in domains such as cybersecurity.
A system crosses into the highest “critical” category if it demonstrates the capacity to independently detect and create working zero-day vulnerabilities of varying severity levels against many well-protected real-world systems.
Alternatively, it qualifies if it can independently design and carry out complete, original attack plans against secure targets based solely on a broad objective.
Earlier systems from the company, including one designated GPT-5.6-Sol, were rated at a lower “high” level rather than critical.
OpenAI emphasized that Astra remains unreleased and played no role in a separate earlier security incident involving the Hugging Face platform.
The company stressed the importance of openness with the public and the broader security community regarding this possible advancement in capabilities.
Cybersecurity landscapes are evolving quickly, with AI systems offering dual potential: bolstering defenses while also enabling faster and larger-scale attacks.
In response, OpenAI has intensified testing of its protective measures to ensure they match the possible new capability level.
Specific actions include introducing tighter controls for more powerful models, such as isolated evaluation setups, limited network and tool permissions, stronger encryption and safeguards for model parameters, expanded monitoring systems, and execution in contained environments.
Internal work on Astra that does not currently satisfy these elevated standards has been put on hold.
Additionally, continuous oversight has been applied to all agent-style uses of the model, covering both training and testing phases; this oversight examines the system’s reasoning steps and can halt concerning activity.
The organization plans to collaborate with government bodies and chosen AI safety groups for further independent examination of the model’s abilities.
It will also share suggested security practices with external partners conducting higher-risk evaluations.
OpenAI noted that the same framework previously guided its approach when models neared elevated thresholds in biological domains, prompting expanded testing, external consultation, and additional protections.
The company expressed confidence that highly capable cyber-oriented systems can ultimately assist defenders in spotting and fixing weaknesses ahead of potential adversaries.
It reaffirmed its commitment to partnering with governments, safety institutes, and civil society so that advanced models like Astra and successors are introduced carefully and in ways that serve humanity overall.
No specific release timeline for Astra has been provided, and the current measures may extend the development period. This step marks a notable instance of a leading AI laboratory publicly adjusting its own research pace in response to cyber-related concerns identified through its internal processes.