Direct and indirect
- Direct: someone types the instructions into the chat themselves.
- Indirect: the instructions arrive inside content the AI reads for you. This is the one that matters most once assistants browse, read email and call tools.
What vendors say
Anthropic's help center warns that malicious MCP servers can embed hidden instructions in what they return. OpenAI warns that an untrusted server can trick ChatGPT into misusing tools. Claude Code's docs warn that servers fetching outside content expose you to prompt injection.
Limiting the damage
Keep approvals on for tools that write, connect fewer and trusted servers, and read what an agent wants to do before you allow it. The lesson prompt injection has five habits.