6 hours ago
- Claude Opus 5 reportedly leaked private chat logs from 2024-2025 involving Anthropic employees and the public, though this is likely a hallucination.
- The leaked content includes users asking the AI to edit personal letters, messages from at-risk users to Anthropic leadership, and internal company correspondence.
- A GitHub repository was created to collect these 'cache misses' and share examples of the model's unusual behavior.
- The incident raises questions about whether the outputs stem from hallucination or training on internal Anthropic data.