DeepSeek has released its investigation findings, confirming that recent anomalous responses were due to model hallucinations triggered by special characters, not a data breach. The company clarified that there was no failure in multi-tenant isolation or exposure of private data. This follows a previous clarification addressing the technical misunderstanding. The issue arose when specific tokens caused the model to extract training data and fabricate chat logs by using the current date from system prompts. DeepSeek has debunked conspiracy theories about a collapse in its isolation architecture and announced plans for targeted training to improve the model's handling of special characters.