OpenAI Curbs Cybersecurity Features on New Astra Model Over Exploit Risks

OpenAI is limiting the advanced cybersecurity functions of its upcoming Astra artificial intelligence model after internal evaluations revealed it could autonomously identify and create novel cyberattacks, the company announced Tuesday.

The decision was made after Astra reached the "critical cybersecurity threshold" under OpenAI's Preparedness Framework, a first for any of its models. This designation signifies the model's ability to discover and weaponize previously unknown security flaws without direct human guidance.

'Critical' Risk Threshold Triggered

During internal testing, OpenAI found that the Astra model was capable of identifying and developing zero-day exploits, which are attacks targeting security vulnerabilities that are not yet known to software vendors or cybersecurity professionals.

Specifically, the model discovered and utilized two distinct zero-day vulnerabilities as part of a successful exploit chain against well-protected systems. Following these findings, OpenAI stated it paused some internal development on Astra in August to implement more stringent safety measures.

Phased Rollout and New Safeguards

To mitigate the risk of misuse, OpenAI will roll out Astra's capabilities in stages. Initial access to the model's advanced cybersecurity tasks will be confined to a small, controlled group of testers.

Ad

Subsequently, a larger group of approved users will gain access for defensive cybersecurity purposes through the company's Daybreak Blue program. OpenAI has also enhanced its safety protocols, including continuous monitoring for unauthorized behavior and systems that can automatically halt potentially malicious activity.

As a result of these new guardrails, the company reported that Astra now refuses 91.5% of malicious cyber-related requests, a significant improvement from the 59% refusal rate of a model identified as GPT-5.6 Sol. OpenAI noted that on August 28, it restarted a major reinforcement learning process that had been paused pending the implementation of these new safety requirements.

Implications for AI Safety

The development highlights the dual-use nature of powerful AI models, which can be leveraged for both sophisticated defense and offense in the digital realm. OpenAI acknowledged that its enhanced safety systems may occasionally flag legitimate activity, potentially slowing or stopping user work.

The company plans to release more detailed information about its safety and security testing in the model's official system card upon its public launch.