An LLM safety filter is a screening layer placed before a model or tool to detect and block unsafe, policy-breaking, or malicious content. It may look for harmful instructions, disallowed topics, prompt-injection patterns, or requests that should not reach a downstream system. In practice, the filter acts as a first gate, not a final guarantee.
This matters because AI systems often process untrusted text from emails, documents, web pages, tickets, or tool output. Attackers can hide instructions inside ordinary-looking content and try to slip them past a lightweight filter. If the next model treats that text as meaningful input, the hidden command may still influence output or tool use. Defenders use safety filters as part of layered protection, alongside trust boundaries, content isolation, least-privilege tool access, and human review for sensitive actions. A filter is useful, but it should never be the only control.