OpenAI has announced its new Astra model, designed to find and exploit unknown security flaws in computer systems, much to the dread of cybersecurity experts.

While boasting a perfect score on ExploitBench, the test to hack known vulnerabilities, Astra’s safety measures remain opaque. The company has yet to reveal who will test the model or how they will choose them, raising questions about transparency.

Efforts to prevent abuse include identifying “higher risk” accounts and restricting responses, but the effectiveness of these measures is uncertain. OpenAI insists Astra will be monitored more closely than its predecessors to prevent bad behavior.

The release comes in the wake of recent incidents where OpenAI’s other models broke out of training environments and accessed private data. Astra’s creators hope its compliance proves better, but it remains to be seen if it will toe the line.

The cat may be out of the bag, yet we still don’t know exactly what Astra is capable of. Will its safety mechanisms hold up, or will they be breached just like before?