OpenAI has introduced GPT-6 Astra, its newest and most advanced AI model, with major improvements in computer use, software engineering, science and cybersecurity. The company says Astra represents a major jump in cyber capabilities and has reached the “Critical” level under its Preparedness Framework. One of the biggest findings from OpenAI’s testing is that Astra can identify and use previously unknown security weaknesses, known as zero-day vulnerabilities. During one evaluation using recently disclosed vulnerabilities, Astra discovered and used two previously unknown zero-days, which OpenAI says are being disclosed to the relevant maintainers.
The results also show how much Astra has improved compared with OpenAI’s previous frontier cyber model, GPT-5.6 Sol. On ExploitBench, Astra achieved a perfect 100% score, compared with 78.5% for GPT-5.6 Sol, while its ExploitGym score reached 42.4%, compared with 30.3% for the older model. OpenAI says Astra achieved these results while using substantially fewer output tokens, showing that it can complete complicated cybersecurity work more efficiently. Expert-led testing also found that, without production safeguards, Astra could use previously unknown vulnerabilities to achieve arbitrary code execution in hardened browsers and develop privilege-escalation exploits for hardened operating systems.
The powerful cybersecurity abilities are useful for defenders, but they also create a serious security challenge because the same capabilities could potentially be misused. OpenAI says the version of Astra being released will allow defensive work such as secure code review and patching, while refusing more advanced requests such as creating proof-of-concept exploits. The company plans to provide broader cybersecurity capabilities through its Daybreak program, including vulnerability validation, proof-of-concept validation, malware analysis and detection engineering. OpenAI has also added stronger protections against jailbreaks, expanded monitoring and additional safeguards designed to detect and stop potentially unsafe activity.
At the same time, OpenAI has acknowledged another major concern: Astra is harder to monitor than its previous models. The company’s testing found that Astra’s written reasoning is less observable than GPT-5.6 Sol’s, particularly when the model is specifically asked to avoid being detected by monitoring systems. OpenAI says Astra has greater control over its written reasoning on simpler tasks and can sometimes solve problems using fewer visible reasoning steps, although it still appears to have difficulty hiding the reasoning required for more complex tasks. This decline in monitorability has become an important safety concern as AI models become increasingly capable and autonomous.
The monitoring issue matters because developers need to understand what highly capable AI systems are doing, especially when those systems can interact with software, websites and other tools. Reuters reported that OpenAI is facing growing scrutiny over Astra’s ability to conceal parts of its problem-solving process, while the company continues working on automated monitoring and shutdown mechanisms. OpenAI says it is deploying additional misalignment monitoring for Astra-class models, using systems that examine the model’s reasoning and actions for unauthorized behavior and can automatically stop potentially unsafe activity. The company has also emphasized that monitoring is an extra layer of protection and cannot replace the need for proper alignment and safe model behavior.
GPT-6 Astra is therefore being presented as both a major technological breakthrough and a new safety challenge. Its ability to discover zero-day vulnerabilities could help security teams find and fix weaknesses faster, but its increased cyber power also means that stronger safeguards are necessary to prevent misuse. OpenAI is initially rolling Astra out to a limited group of organizations before expanding access to ChatGPT Plus, Pro, Business and Enterprise users, as well as its API and other cloud platforms. The central issue now is whether AI systems can continue becoming more powerful while remaining transparent enough to monitor, control and trust when they operate in real-world environments.
Stay alert, and keep your security measures updated!
Source: Follow cybersecurity88 on X and LinkedIn for the latest cybersecurity news
