News
OpenAI Suspends Next-Generation AI Model Astra’s Cyber Attack Capabilities Are Too Dangerous
2 min read
Source: newmobilelife.com
OpenAI said today that it is "pausing" activities involving its upcoming AI model Astra because its cyber attack capabilities may be too dangerous. OpenAI pointed out that the latest internal evaluation showed that Astra has made "significant progress" in agent coding and network security, so it cannot be ruled out that it has "critical network attack capabilities"; OpenAI's past AI models, including GPT-5.6 Sol, were only marked as "high". Triggering strict guidance Astra triggered the stricter guidance in the OpenAI "Readiness Framework". The guidance calls for caution when developing cutting-edge AI capabilities that pose serious risks of harm; the cybersecurity portion of the framework also states that OpenAI will implement additional safeguards for AI models that "create new risks of large-scale cyberattacks and vulnerability exploitation." The "critical" threshold that Astra may have reached is defined as identifying and developing effective zero-day vulnerabilities across all severity levels that can be applied to multiple hardened real-world critical systems without human intervention, or conceiving and executing novel end-to-end cyber attack strategies. Strengthening protection before deployment OpenAI said that safeguards and security controls will be strengthened before Astra is deployed, including limiting work involving the AI model until new safeguards are in place. The company plans to use an isolated testing environment with restricted network and tool access, and add sandbox execution and more monitoring functions; it will also cooperate with relevant government agencies and AI safety organizations to test Astra. OpenAI said: "We are committed to working with governments, security institutes and civil society to ensure that the cutting-edge capabilities of Astra and subsequent similar AI models are deployed in a responsible and widespread manner to benefit all mankind." Security testing background Astra has not been officially announced, but OpenAI recently shared the details of the next generation of major AI models in an article about mathematical progress; Astra costs about $2,000 (per Sol API rate calculation) solves the math and theory