On September 3, OpenAI unveiled GPT-6 Astra, a new artificial intelligence model that completed the video game Portal entirely on its own. The AI played the game without human guidance, using only screenshots and issuing its own control commands to navigate through the puzzles. The task required 3,336 calls to the model, costing $571.18. While the actual gameplay lasted about two hours, the AI spent much of the day analyzing each scene and deciding on its next move, making the total time span nearly 24 hours. OpenAI, based in San Francisco, is a U.S.-based public benefit corporation focused on advancing AI. The company is known for its series of large language models, called GPT, including the widely used ChatGPT, which became one of the most visited websites in the world by 2026. GPT-6 Astra, according to OpenAI, represents a major leap in AI capabilities, promising to boost entrepreneurship, creativity, and scientific progress. The model will be rolled out in stages, with priority access given to companies participating in the "Daybreak" cybersecurity initiative. Despite the impressive feat, the achievement sparked mixed reactions. Some praised the AI’s ability to solve a complex puzzle without human input, while others questioned the novelty of the accomplishment. Portal, a puzzle game first released in 2007, has been extensively analyzed online, with detailed guides available for free. Critics argue that the AI may have simply replicated solutions it encountered during its training, rather than solving the game in a truly novel way. Additionally, the game’s structure—each room having a single correct solution—makes it a closed problem, unlike more open-ended tasks. A deeper concern for AI security experts lies in the "deep recurrent" nature of GPT-6 Astra, which allows the model to perform internal reasoning loops without leaving clear traces of its decision-making process. This lack of transparency raises alarms about the difficulty of monitoring and understanding how the AI operates, especially in critical applications. While the ability to detect and exploit "zero-day" vulnerabilities is a concern, the broader issue is the reduced visibility into the AI’s behavior, which could complicate efforts to ensure its safety and reliability.