New release of LLM adds support for reasoning traces, OpenAI Responses, server-side tools, and smarter logging

OpenAI is temporarily halting development on its upcoming AI model, Astra, after internal testing revealed the potential model may have reached a cybersecurity limit that no previous OpenAI model has ever encountered. While the company has not set an official release date for Astra, the setback could delay its release. OpenAI stated that the model could possess "critical cyber capabilities," and paused activities that do not meet heightened security requirements while expanding testing and tightening controls around the model.

Astra is now being treated as a Critical cybersecurity model, requiring additional safeguards during development. This classification is based on the model's ability to identify and develop functional zero-day activities in various hardened, real-world critical systems without human involvement, or by performing novel end-to-end attacks against hardened targets after receiving only a high-level goal.

OpenAI is testing Astra in isolated environments with stricter network and tool limits, as well as enhancing protection around the model itself and adding supervision to stop it from performing unsafe behavior. Although the company has not confirmed whether Astra will be accessible through ChatGPT, Codex, or its API, it has previously offered Trusted Access for Cyber, a program that provides approved security professionals with tools not available to everyone.

Astra's capabilities might be limited to researchers and organizations willing to accept more oversight, with developers required to prove their identity and explain the intended use of the model before gaining access to its most powerful capabilities.