OpenAI issued a security announcement on August 7, local time, stating that internal assessments showed that its next-generation model “Astra” has made significant progress in intelligent programming and cybersecurity capabilities, and the company is currently unable to rule out the possibility that this model will reach a “critical level” level of cybersecurity capabilities. According to OpenAI, the relevant assessment results have triggered security response measures in its “Preparation Framework”. According to OpenAI's definition, the model has reached a “critical level” of network security capabilities, which means it can discover and develop various effective zero-day vulnerabilities for multiple real critical systems with enhanced protection without human intervention, or formulate and execute novel end-to-end cyber attack strategies for high-protection targets only based on high-level attack targets. OpenAI said Astra has not been used in previous security incidents involving Hugging Face. In response to this model, the company has strengthened security measures, including using isolated test environments, restricting access to networks and tools, strengthening model weight protection and encryption, and increasing monitoring and inspection capabilities.

OpenAI issued a security announcement on August 7, local time, stating that internal assessments showed that its next-generation model “Astra” has made significant progress in intelligent programming and cybersecurity capabilities, and the company is currently unable to rule out the possibility that this model will reach a “critical level” level of cybersecurity capabilities. According to OpenAI, the relevant assessment results have triggered security response measures in its “Preparation Framework”. According to OpenAI's definition, the model has reached a “critical level” of network security capabilities, which means it can discover and develop various effective zero-day vulnerabilities for multiple real critical systems with enhanced protection without human intervention, or formulate and execute novel end-to-end cyber attack strategies for high-protection targets only based on high-level attack targets. OpenAI said Astra has not been used in previous security incidents involving Hugging Face. In response to this model, the company has strengthened security measures, including using isolated test environments, restricting access to networks and tools, strengthening model weight protection and encryption, and increasing monitoring and inspection capabilities.