OpenAI launches GPT-6 Astra with critical cyber capability

OpenAI president Greg Brockman has declared that the era of artificial general intelligence (AGI) has begun with the launch of their latest model, GPT-6 Astra. This model is being hailed as a "generational leap in capability," capable of performing more tasks faster than ever before. Astra's cybersecurity capabilities are so advanced that they meet OpenAI's "Critical" threshold. However, the company assures that there won't be any more agents going rogue, probably.

GPT-6 Astra is positioned as the world's most intelligent and aligned model, excelling in computer use, browsing, software engineering, cybersecurity, science, and professional work. The model scored an impressive 99.9% on ARC-AGI-3 and 97.6% on FrontierMath Tier 4. It also outperformed its predecessor, Sol, in terms of speed, completing tasks in around 40 minutes compared to Sol's 75 minutes. Astra's Codex setup was 1.9 times faster in Mind2Web tasks.

Brockman believes the significance of Astra goes beyond just impressive benchmarks. He suggested that it might mark the arrival of AGI, stating, "I think it might be about this model." The company's announcement focused heavily on Astra's cybersecurity capabilities, with the model scoring 100% on ExploitBench, discovering two zero-days during evaluation, and achieving arbitrary code execution against hardened browsers.

The public version of Astra will refuse requests for proof-of-concept exploits, with broader access reserved for vetted defenders through Daybreak.

Despite Astra's impressive performance, it didn't score as highly as some other models on certain tests, such as Humanity's Last Exam and Artificial Analysis Intelligence Index. The milestone that will likely draw the most attention is its cybersecurity performance. Under OpenAI's Preparedness Framework, "Critical" means the ability to find unknown vulnerabilities and exploit well-protected systems without step-by-step human guidance.

Astra achieved a perfect score on ExploitBench, discovered two zero-days during evaluation, and executed arbitrary code against hardened browsers. The public version of Astra will have limitations on proof-of-concept exploits, while broader access is reserved for vetted defenders through Daybreak.