| |
OpenAI is addressing the safety risk of adversarial distillation, where bad actors can extract capabilities from frontier AI models and redeploy them without appropriate safeguards. The company has implemented stronger protections across detection, training, classifiers, and API-level defenses, particularly focusing on preventing extraction of models' internal reasoning processes (chain-of-thought). In testing GPT-6.2 Astra, OpenAI reduced the success rate of attacks aimed at recovering protected reasoning from 54% to 0.2%.
Read Full Article →
← More Tech news