OpenAI Launches 'GPT-6 Astra,' Boosting Agent and Cybersecurity Performance

OpenAI on the 3rd (local time) unveiled and launched its next-generation AI model, "GPT-6 Astra." OpenAI announced that it will first make GPT-6 Astra available to some enterprise customers participating in the cybersecurity program "Daybreak." It also said it plans to gradually expand the safety-equipped version to paid ChatGPT users and the API.

"GPT-6 Astra" is a model that strengthens agent capabilities and cybersecurity performance. The model aims to autonomously handle complex tasks, focusing on expanding beyond the existing generative AI role centered on simple question-and-answer and information summarization.

Astra is designed to handle end-to-end multi-step work such as computer use, browsing, software engineering, and science. It also enhances the ability to independently judge delegated tasks and carry them out using tools.

A key feature of the new model is its changed approach to tool use. By supporting "Async Tool Calling," it allows other inference and tool calls to proceed in parallel while handling external tool tasks, and it was designed with "Mid-turn Steering" so that user requirement changes or additional instructions during a task can also be reflected.

OpenAI said it disclosed examples showing that these changes lead to shorter real-world task completion times. Astra completed a search for pet boarding services in 5 minutes 27 seconds, while the human time required for that search was about 30 minutes. It also finished a job search in 2 minutes 51 seconds, while the required time for that search was about 5 hours.

Greg Brockman, OpenAI president, said the new model could eventually be regarded as an early stage of AGI. He said there is still much to improve, but explained that there are important areas that have qualitatively improved significantly. He added that these changes mean changes in the kinds of tasks people can delegate to AI and changes in the capabilities AI can give people.

OpenAI emphasized expanded context for handling large-scale tasks, citing Astra's support range as a 1.05 million-token context window and a maximum output of 128,000 tokens. API pricing is USD 10 per 1 million tokens for input and USD 50 per 1 million tokens for output.

OpenAI also highlighted improved cybersecurity capabilities alongside the expanded scope of task execution. Cybersecurity performance has improved from the previous model, and in the OpenAI Preparedness Framework evaluation, its cybersecurity capability reached the Critical stage for the first time. With the proper tools and access privileges, it can detect unknown vulnerabilities and also has the conditional ability to develop ways to exploit vulnerabilities without step-by-step human instruction.

In response, OpenAI said it will strengthen refusal of harmful requests, tighten access controls, and reinforce systems for monitoring unusual behavior. It also set out a usage policy in which high cybersecurity capabilities will be used for defensive purposes such as vulnerability detection and patch development, while access that could enable offensive misuse will be restricted.

Along with Astra's launch, OpenAI said it will expand support for cyber defense and operate "Daybreak for Frontier Defenders." The program will support security researchers and organizations using AI, and plans to help with vulnerability discovery, validation, and patch development by leveraging Astra's cybersecurity capabilities.

The timing of this release coincides with a phase of increasing autonomy in AI agents, and the importance of safety controls has also grown as autonomy expands. OpenAI experienced a test AI agent incident last July, involving departure from an isolated development environment and access to external systems. It later took corrective measures, including strengthening development environment isolation and access controls, and said that although Astra was not the model used in that incident, the issues identified at the time were reflected in the safety system for subsequent models.

OpenAI said it will begin offering Astra to a limited number of enterprise customers and expand access in stages. As the scope of agent task handling and tool use expands, OpenAI said it will continue strengthening safety systems for task permissions and behavioral controls alongside model performance.

Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=406524

References

• https://www.epnc.co.kr/news/articleView.html?idxno=406524