Common attack paths
- Malicious user input in chat interfaces
- Injected instructions in retrieved documents
- Tool output contamination in agent workflows
- Multi-step prompt chaining without policy gates
Article
Prompt Injection Explained
Prompt injection tricks an LLM into ignoring intended rules and following attacker instructions from user input, documents, or external tools.
Defensive pattern
- Strict system-policy hierarchy
- Input sanitation and instruction detection
- Output validation before actions execute
- Human approval for high-risk operations