OpenAI has revealed new details about the upcoming release of its language model Astra, which has reached a critical threshold in cybersecurity. The system demonstrates the ability to independently detect unknown bugs in computer networks and exploit them for hacking without human intervention.
OpenAI has revealed new details about the upcoming release of its language model Astra, which has reached a critical threshold in cybersecurity. The system demonstrates the ability to independently detect unknown bugs in computer networks and exploit them for hacking without human intervention.
The developers announced this on the official OpenAI blog, outlining the details of the tests and planned access restrictions.
According to the company, Astra received the maximum score in the ExploitBench test, which measures the ability of multi-modal models to break into protected systems. In a modified test by OpenAI engineers, the model was able to independently find and successfully exploit two previously unknown zero-day vulnerabilities.
«We plan to make Astra available soon, but access to its cutting-edge cybersecurity capabilities will be more limited,» OpenAI said in a post.
In early August , OpenAI announced that it was suspending development of certain elements of Astra. An internal audit revealed a significant leap in its coding and cybersecurity capabilities — details that were powerful enough to raise concerns about the security of the development.
To prevent abuse, the company is implementing additional chain-of-thought monitoring to detect malicious behavior, and will block specific requests from high-risk accounts. The model is also being further tested for attempts to «escape» from the isolated sandbox environment.
Despite the developers’ assurances, independent experts have expressed concern about the lack of external auditing. Former OpenAI employee Yona Shavit, who now works at the OpenAI Foundation, noted on social media that Astra’s restraint during security tests could indicate that the model was «aware» of the fact that it was being tested or was trying to adjust its responses to the researchers’ expectations. At the moment, OpenAI does not disclose the composition of the group of testers or specify whether it is cooperating with government regulators to verify security.
A similar level of threat from autonomous exploitation was previously described by Anthropic when it announced its Mythos model. OpenAI’s current preparations for the launch of Astra come amid industry concerns over previous incidents where test AI agents gained unauthorized access to private data on the Hugging Face platform .
