Claude hallucinates internal tools, falsely accuses user of prompt injection

View on Reddit

User reports Claude hallucinated its own tool definitions and accused them of a prompt injection attack. The incident involved leaked JSON definitions from backend tool-calling layers.