An AI agent's persistent memory was poisoned with hallucinated facts within just 48 hours of deployment, causing it to confidently operate on completely false information.

A solo developer built and deployed a full-stack LLM platform (3 API integrations, real-time streaming, React/Express/TypeScript) almost entirely using Cursor + Codex. The tool excelled at scaffolding and pattern replication—until API hallucinations, scope creep, race conditions, and silent failures nearly killed the project in production.

A developer asked an LLM for help with a library API and was given a method name that didn't exist. Googling revealed only one other result—a GitHub issue where someone else had been told the same fictional method by another LLM.