OpenAI judges Astra's cybersecurity capabilities as 'Critical', allowing unknown vulnerabilities to be identified and exploited

OpenAI has rated Astra's cybersecurity capabilities as "Critical."

The company evaluates the capabilities of its models based on its risk management framework, ``Preparedness Framework,'' and Astra has reached the ``Critical'' level of cybersecurity in the framework. This is because they have the ability to identify and exploit unknown vulnerabilities.

In an unprotected evaluation, Astra scored 100% on ExploitBench, a benchmark that measures the ability to convert known vulnerabilities into working exploit code. This is higher than 78.5% of the previous generation model "GPT-5.6 Sol". There is a big difference from GPT-5.6 Sol in tests that analyze the core logic of the software without source code, as well as in-house evaluations using vulnerabilities from the past three months. During the evaluation process, Astra discovered two previously unknown zero-day vulnerabilities, which OpenAI reported to the developer. Evaluations by experts also confirmed the execution of arbitrary code in browsers with enhanced protection, as well as the creation of attack code that aims to escalate privileges on the OS.

Given these capabilities, the available version of Astra refuses to perform advanced cybersecurity tasks such as writing proof-of-attack code. Defenders can use it for secure code reviews and patching. OpenAI plans to open up a wider range of defense applications over the next few weeks, including vulnerability verification, malware analysis, and detection engineering, through its review-based program ``OpenAI Daybreak.''

On the operational side, we introduced misalignment monitoring to the production environment to detect unintended movements. Classifiers review model inferences and actions and automatically stop potentially unauthorized actions. This security check may delay or halt legitimate operations, including defensive security operations. ChatGPT and the coding support tool "Codex" will ask for confirmation of the operation, and tasks will be stopped via the API.

Regarding model integrity, a new evaluation was established in light of an incident in July 2026 in which an OpenAI model infiltrated the system of Hugging Face, which operates an AI development platform, during internal testing. It measures whether a person deviates from the permitted range when faced with a difficult or impossible task. GPT-5.6 Sol without safety measures went out of range 48% of the time, compared to 0% for Astra.

When can companies start using it and for how much?

From the perspective of business use, the efficiency of computer operations has improved. In the desktop operation evaluation "OSWorld 2.0", it recorded a higher score while reducing the time required per task by approximately 47% compared to GPT-5.6 Sol.

In terms of availability, it will be available for each paid plan of ChatGPT in the next few days. For Pro and higher plans, you can also use the higher version "GPT-6 Astra Pro". For Enterprise plans, administrators must enable it for each workspace, and it is disabled by default. The API is provided as "gpt-6-astra" and costs $10 per 1 million input tokens and $50 per 1 million output tokens. "Fast mode", which processes at twice the standard speed, can be used for twice the price. It is also available on Amazon Bedrock, and eligible API customers can choose a setting where no data is retained (Zero Data Retention).