OpenAI Classifies GPT-6 as Highest Cybersecurity Risk

OpenAI has classified its new model, GPT-6 Astra, as 「Critical」—the highest risk category for cybersecurity—under its internal safety evaluation standard known as the 「Preparedness Framework」. This is the first time any of the company's models has reached this classification level.

The Preparedness Framework is OpenAI's proprietary safety standard established to systematically evaluate potential risks before releasing powerful AI models to the public. It assesses risks from multiple perspectives, including potential misuse for chemical and biological weapons and applicability to cyberattacks, assigning risk levels to each category. The fact that GPT-6 Astra has reached the highest level of 「Critical」 indicates that OpenAI itself has determined the model possesses capabilities in the cybersecurity domain that are in a different category from existing models.

Testing by an expert team revealed that GPT-6 Astra discovered previously unknown vulnerabilities in both browsers and operating system (OS) kernels. Furthermore, it successfully created what is called an 「exploit」—attack code that can actually leverage these vulnerabilities. This demonstrates not merely the identification of vulnerabilities, but the ability to independently construct tools usable in actual attacks, making it a significant finding in the test results.

Meanwhile, the same system card (the model's safety evaluation report) also documented another concern: a substantial decline in the monitorability of 「chain-of-thought」reasoning. Chain-of-thought refers to the visualization of inference steps that an AI takes to arrive at an answer, serving as a crucial reference point for researchers and safety personnel to verify the model's intent and decision-making process. A reduction in this transparency means that the model's internal operations become harder to discern from the outside.

This situation is deeply linked to a long-standing challenge in AI safety. The 「capability-risk tradeoff」—the notion that higher-performing models entail higher potential risks—has been recognized before, but this case is distinctive in that the enhancement of capabilities and the decline in monitorability occur simultaneously. The combination of increased capability for applying to cyberattacks while making internal processes less visible raises new questions about how safety management should operate.

As attention to the safety evaluation of 「frontier models」has grown across the AI industry, OpenAI's application of the highest risk classification to its own model can be regarded, in terms of evaluation process transparency, as a measure of progress. However, it simultaneously raises a critical question: under what conditions should such a model be released and operated? Going forward, the key focus of attention will be what safety measures and access restrictions are implemented for models classified as Critical.