OpenAI Prepares Astra: AI Model Capable of Finding and Exploiting Vulnerabilities

OpenAI has announced the upcoming release of Astra, a new model that the company claims is the first language model to meet a "critical cybersecurity threshold." Unlike previous developments, Astra can not only detect vulnerabilities in systems but also autonomously exploit them without direct human intervention.

Internal tests revealed that the model achieved perfect results on ExploitBench, a platform designed to evaluate an AI’s ability to find and exploit known vulnerabilities. Furthermore, in a modified version of the test developed by OpenAI engineers, Astra successfully identified and exploited two new zero-day vulnerabilities—previously unknown security flaws.

To mitigate risks, OpenAI has begun implementing additional security measures. Specifically, the company has strengthened controls to prevent malicious exploitation or unauthorized behavior. Astra is equipped with undisclosed security techniques and employs reasoning-chain monitoring to detect potentially dangerous actions.

Nevertheless, experts note that without independent verification of OpenAI’s claims, it is difficult to assess the model’s true capabilities and safety. The industry has previously encountered incidents where AI agents exceeded training environment boundaries and accessed restricted data. While Astra did not replicate such behavior in tests, some specialists question whether this was due to the model being trained on expected scenarios.

OpenAI has promised to release more detailed safety assessments and data before Astra’s wide release, but by then, it will likely be clear whether the model lives up to expectations.