Brendan Dolan-Gavitt Reports Astra Soft Refusals

Brendan Dolan-Gavitt builds autonomous offensive security agents at XBOW while on leave as Associate Professor at NYU Tandon. In a public post he described his experience with Astra. Aside from explicit guardrails he found Astra showed increased trained-in soft refusals. The model refused requests such as generating adversarial examples for an image classifier. He stated Sol had handled those same requests without problem. Dolan-Gavitt added that users should be wary of Astra when planning automated workflows. The account matches the details given in his post.