OpenAI unveiled a new generation model, GPT-6 Astra, on Thursday (September 3). The company stated that Astra’s cybersecurity capabilities have significantly improved compared to GPT-5.6 Sol, making it the first model to meet the “Critical” cybersecurity threshold set in its “Preparedness Framework.”
On the same day, OpenAI announced the launch of the “Daybreak for Frontline Defenders” initiative, committing to provide $1 billion worth of subsidies, training, and technical support to assist U.S. public institutions such as water, electricity grids, and local governments in strengthening their network defenses.
Astra is the latest flagship model following GPT-5.6 Sol. Unlike AI primarily answering through dialogue, Astra further enhances its “hands-on” capabilities, enabling it to operate computers, use online tools, complete multi-step tasks consecutively, such as assisting with tax filing, game development, creating architectural renderings, organizing legal documents, and finding rental properties.
According to Reuters, OpenAI’s CEO Greg Brockman stated that the new model expands the scope of tasks that can be assigned to AI. He also believes that Astra may signal the arrival of the era of Artificial General Intelligence (AGI).
OpenAI also released two contrasting sets of numbers: when searching for appropriate cat caretakers online, Astra took 5 minutes and 27 seconds, while humans took about 30 minutes for the same task; in another online job search task, Astra took 2 minutes and 51 seconds, whereas humans needed about 5 hours to complete. These numbers are from tests released by OpenAI and are not from independent third-party evaluations.
One significant enhancement of Astra is in the field of cybersecurity. OpenAI states that Astra can quickly detect security vulnerabilities in corporate systems and becomes the first model to achieve the “Critical” level of cybersecurity capability.
According to OpenAI’s definition, this means that with the appropriate tools and permissions, the model can rely less on step-by-step human guidance, search for previously unknown vulnerabilities in multiple well-protected systems, and develop new exploitation methods.
This capability could also be subject to abuse. OpenAI mentioned that the version released to ordinary users would limit higher-risk cybersecurity tasks; the more comprehensive cybersecurity capabilities will be provided through controlled programs like Daybreak to verified defenders. The system may slow down, pause, or terminate tasks if unauthorized behavior is detected.
OpenAI also stated that as the model’s capabilities improve, monitoring and assessing the reasoning process through the model become increasingly challenging.
Chief Scientist of OpenAI, Jakub Pachocki, stated that as the model’s capabilities advance, accurately determining what they can ultimately do also becomes more difficult.
On the day of Astra’s release, OpenAI announced the aforementioned $1 billion commitment. This is not a cash donation of $1 billion, but rather $1 billion worth of Daybreak subsidies, training, technical support, and related partnerships. The company plans to provide these resources over the next approximately six months.
The initial recipients will be U.S. institutions with relatively limited network defense resources but crucial to public services, including water and wastewater treatment systems, electricity grids, state and local governments, community and regional banks, non-profit organizations, and more.
OpenAI also announced a new pilot collaboration with the Multi-State Information Sharing and Analysis Center (MS-ISAC). This center provides network threat intelligence, incident response, and information sharing services to thousands of U.S. public sector organizations, including utilities, public hospitals, K-12 schools, and law enforcement agencies.
In the coming weeks, OpenAI plans to expand the Daybreak project to partner countries of the United States.
The company anticipates that as AI model capabilities improve, using AI to assist in launching network attacks will become more widespread and complex in the coming months. Daybreak will provide defense personnel with AI tools to discover vulnerabilities, verify security issues, and expedite patching.
Previously, over 100 companies including OpenAI, Anthropic, Microsoft, Alphabet, and Amazon collectively warned that the time for defenders to strengthen their defenses against AI-driven network attacks is shrinking before these attacks become more prevalent.
Astra was first made available to select institutions on September 3. OpenAI stated that in the following days, it will gradually roll out to ChatGPT Plus, Pro, Business, and Enterprise users; enterprises and developers can also access Astra through API interfaces and Amazon Web Services (AWS).