A cybersecurity vulnerability where an attacker manipulates an LLM's inputs to bypass its safety guardrails and execute unauthorized commands.
Prompt Injection is comparable to SQL Injection: in both, untrusted input ends up treated as instructions. If an enterprise connects an LLM to internal databases and APIs, a malicious user can craft input that overrides the system prompt and directs the AI to delete records or exfiltrate data. Mitigating this threat requires strong Control Boundaries and independent verification layers.
See also Control Boundaries, Air-Gapped Environment, Role-Based Access Control and Zero Trust Architecture.