CoinWorld reports that a user on GitHub has reported that while using Anthropic's Claude Code, server IPs, usernames, and plaintext root passwords that do not belong to them appeared in the conversation context, causing the AI to automatically connect to other servers and modify databases. Community technical analysts believe the root cause is a failure in the isolation mechanism of the large model's prompt prefix cache, leading to collisions between cache keys of different users. GitHub has now labeled the issue with the tag "area:security," and all parties are awaiting Anthropic's verification conclusion.