Security firm Xint discovered a method using OpenAI's compaction feature to extract a model's encrypted reasoning into plain text.The technique bypassed previous replay blocks by running compaction on the same model and passing the compressed item to cheaper models.Xint confirmed the issue, which OpenAI fixed on October 23, awarding researchers $6,000.https://en.hacks.gr/ereynites-vrikan-tropo-na-diavasoyn-tis-kryfes-skepseis-tis-technitis-noimosynis-tis-openai/#OpenAI #ArtificialIntelligence #Privacy #MachineLearning
Security firm Xint discovered a method using OpenAI's compaction feature to extract a model's encrypted reasoning into p...
