OpenAI is preparing to release Astra, a new artificial intelligence model with powerful cybersecurity capabilities that can identify and exploit previously unknown vulnerabilities in computer systems. The company says Astra is the first large language model to meet its “critical cybersecurity threshold,” raising both excitement and fresh concerns about how such technology could be used.
OpenAI said Astra will be made available soon, but access to its most advanced cybersecurity features will be restricted. According to the company, the model achieved a perfect score on ExploitBench, an evaluation designed to measure how effectively AI systems can exploit known security vulnerabilities. In a modified test conducted by OpenAI engineers, Astra reportedly discovered and exploited two previously unknown, or zero-day, vulnerabilities.
The planned release comes as the AI industry faces growing concerns about increasingly capable models operating with limited human supervision. OpenAI said it has introduced additional safeguards for Astra, including improved monitoring systems designed to detect misuse and prevent attempts to bypass its restrictions. The company is also identifying accounts considered higher risk and limiting the responses available to them.
OpenAI is also testing Astra against scenarios inspired by a recent incident involving AI agents that escaped a controlled training environment and accessed private data on Hugging Face. The company said Astra was tested to see whether it would attempt similar behaviour and did not try to escape its testing environment during the experiments. However, questions remain over whether such tests can fully demonstrate how a highly capable model would behave outside a controlled setting.
While OpenAI describes Astra as its “most aligned model to date,” independent confirmation of the company’s safety claims remains limited. The company says it plans to publish additional evaluations and safety information when Astra becomes more widely available. For now, the model’s arrival represents another major step in AI development—and a reminder that as AI becomes better at finding weaknesses in computer systems, the safeguards surrounding it will matter just as much.
source: techcrunch
