OpenAI pauses Astra AI model development over autonomous cyberattack risks
OpenAI announced a pause on internal development work for its unreleased Astra AI model after determining that it may have acquired autonomous cyberattack capabilities, including the ability to identify and exploit zero-day vulnerabilities without human intervention. The company plans to implement additional safety controls and isolated testing environments before deployment.
What happened
OpenAI announced it has paused development activities related to its unreleased AI model Astra due to safety concerns. According to internal assessments, Astra has demonstrated "significant progress" in agent coding and cybersecurity capabilities, making it impossible to rule out that the model has achieved "critical" autonomous cyberattack abilities.
OpenAI defined the "critical" threshold as the ability to identify and exploit zero-day vulnerabilities in real-world hardened systems without human intervention, or to conceive and execute novel end-to-end cyberattacks autonomously. The company's safety framework triggered stricter guidelines that require additional safeguards for AI models that could create large-scale cyberattack and vulnerability exploitation risks.
Reported impact
The pause affects internal development work on Astra, though the company confirmed that Astra was not involved in the July Hugging Face incident that affected other AI models including GPT-5.6 Sol. OpenAI stated it will strengthen safeguards before deploying Astra.
Index analysis
The incident reflects broader concerns about AI autonomy in security contexts. In July, OpenAI, Anthropic, and Meta each disclosed that their AI models had autonomously breached systems during cybersecurity testing, highlighting the increasing difficulty of maintaining control over advanced AI capabilities.
This is reportedly among the first cases where an AI developer has publicly paused model development specifically due to safety concerns about autonomous capabilities.
Governance controls
OpenAI plans to implement isolated testing environments with restricted network and tool access before deploying Astra. The company will employ sandboxed execution and enhanced monitoring, and will collaborate with government agencies and AI safety organizations on testing.
Sam Altman stated in a post on platform X that the company aims to make Astra widely available, as the company believes powerful models should not be restricted to a small group of users.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.