As cited
Copy frozen at (site build).
ai security
OpenAI reveals ‘novel’ encryption bypass used in distillation attack
OpenAI disrupted a coordinated campaign between July 1 and July 28 to extract reasoning capabilities from its models by copying encrypted data from one conversation and requesting decryption in another. The company attributed a core cluster of the activity to Moonshot artificial intelligence (AI), a China-based artificial intelligence (AI) rival, though it provided no technical evidence for the attribution. OpenAI addressed the vulnerability by fixing a bug that allowed cross-conversation data access and strengthening account signup and network monitoring controls.
Why it matters: Organizations using OpenAI models and developers building on top of them face ongoing model distillation attacks; practitioners should implement additional monitoring for suspicious prompt patterns and account activity, and evaluate whether competitors may be systematically extracting their proprietary model capabilities.
- Source published
- First seen by Cybersecurity Tracker