OpenAI has paused internal development activities for its upcoming Astra AI model after evaluations indicated the system may have reached a "critical" cybersecurity threshold. Under OpenAI's Preparedness Framework, this designation means the model can autonomously identify zero-day exploits and execute end-to-end cyberattacks without human intervention. OpenAI is implementing isolated testing environments and universal monitoring, which CEO Sam Altman confirmed will delay Astra's public launch.
Astra model cybersecurity capabilities
- ▪OpenAI announced on August 7, 2026, that internal evaluations of its upcoming Astra AI model indicated significant advancements in agentic coding and cybersecurity
- ▪OpenAI confirmed that the upcoming Astra AI model was not involved in the recently disclosed security exploit on the Hugging Face platform
- ▪OpenAI stated that it cannot rule out that its upcoming Astra AI model has reached the "critical" cybersecurity capability level under its Preparedness Framework
OpenAI development pause
- ▪A White House official stated that OpenAI voluntarily informed the Biden administration of its plans to delay the release of the Astra AI model
- ▪OpenAI paused internal development activities involving the Astra AI model that do not meet newly strengthened security control requirements
- ▪OpenAI is implementing stricter security controls for the Astra AI model, including isolated testing environments, restricted network and tool access, and universal monitoring
Critical cyber threshold definition
- ▪Under OpenAI's Preparedness Framework, a model reaches the "critical" cybersecurity threshold if it can devise and execute novel end-to-end cyberattack strategies against hardened targets given only a high-level goal
- ▪Under OpenAI's Preparedness Framework, the "critical" cybersecurity threshold is reached when a model can autonomously identify and develop functional zero-day exploits of all severity levels in hardened real-world systems
Altman comments on competitors
- ▪OpenAI CEO Sam Altman confirmed via X that the cybersecurity assessment will delay the launch of the Astra AI model
- ▪OpenAI CEO Sam Altman criticized rival Anthropic on X, stating that keeping powerful AI models restricted to a chosen few is not a good strategy
Preparedness Framework safety protocols
- ▪OpenAI's previous high-end model, GPT-5.6 Sol, only reached the lower "high" cybersecurity threshold during its internal evaluations
- ▪OpenAI's Preparedness Framework, first published in December 2023, outlines specific capability thresholds in categories such as cybersecurity, biological risks, and AI self-improvement that trigger development halts
Story comments
Loading comments…