OpenAI says it plans to publicly release a version of Astra "soon" but will make its advanced cyber capabilities available only to select partners (Wired)

OpenAI has announced plans to release a version of its Astra AI model, but will only make its advanced cybersecurity capabilities accessible to select partners. The company reached the critical cyber threshold for Astra when it demonstrated the ability to independently find and exploit previously unknown vulnerabilities in real-world software.

Following a multi-week pause for additional safety and security controls, OpenAI says it is now confident in releasing Astra broadly in a safe manner. To limit everyday users from accessing Astra's advanced cyber capabilities, the company has implemented a new "misalignment monitor." If a user asks Astra to help find an exploit in a real-world software system, the model is supposed to refuse to answer.

OpenAI has also made the AI model more robust against jailbreaking attempts and tested it to have a higher refusal rate to unsafe queries than previous models. However, OpenAI notes that its misalignment monitor may occasionally flag legitimate activity as potential cyber misuse or unauthorized behavior. Partners in OpenAI's Daybreak program will receive early access to a less restricted version of Astra with enhanced cyber capabilities.

Astra's capabilities include finding novel software vulnerabilities, developing ways to exploit them, and chaining multiple exploits together to breach deeper into target systems. OpenAI's Astra outperformed industry-leading AI models in cybersecurity benchmarks such as ExploitBench.