오픈AI ‘아스트라’, 치명적 사이버 역량 도달…안전장치 강화해 출시

OpenAI revealed that its upcoming artificial intelligence model, 'Astra', has achieved a "critical" level of cyber security capabilities in its latest evaluation. The company plans to release Astra soon, after limiting advanced cyber security abilities. OpenAI announced on its blog that after collecting more evidence and conducting additional assessments, they determined that Astra has reached the critical stage in its cyber security capabilities, which is the first time for any of their models to do so.

This marks the first instance among OpenAI's previously developed models to reach this level of capability. OpenAI promoted Astra's achievement in solving 10 long-standing math problems or making substantial progress in the field last month, but they cannot exclude the possibility that the model has now reached the critical level of cyber capabilities just a day later.

According to OpenAI's internal standards, a model reaches a critical level when it can find undisclosed vulnerabilities in a security-enhanced system without human intervention or when it can devise and execute a new cyber attack strategy from scratch given only high-level goals. The company believes Astra has met the first condition, as it found multiple vulnerabilities in a hardened operating system and used them to gain administrator-level access in a non-administrator state.

Astra also recorded a perfect score in a benchmark for developing attack code based on the vulnerabilities. OpenAI has significantly strengthened safety measures to prevent potential misuse of the model, and implemented a system that monitors the model's reasoning process and actions, automatically halting potentially unauthorized activities.

OpenAI reported that serious harm risk is minimized with these safety measures, and Astra will be released soon to a limited number of users for testing. However, advanced cyber security capabilities will initially be offered to a small group of users.