Claude Code leaks user credentials to other users, accidentally modifies external databases. According to Beating monitoring, a GitHub report shows that Claude Code users encountered foreign server IP addresses, usernames, and plaintext passwords in their AI conversation context. The local Claude Code assistant then automatically connected to these unrelated servers via SSH and executed database write modifications, effectively exposing another user's production database to unauthorized changes. Community analysts suspect that the root cause is a failure in the isolation mechanism of the large language model's prompt prefix cache. If cache key collisions or isolation failures occur between different users, confidential information may be erroneously spliced into another user's conversation, potentially exposing server credentials and source code to unintended parties. The issue has been marked as security-related on GitHub, pending official investigation. #预测世界杯葡萄牙VS西班牙