Security · AI
Firewall LLM
The security guard inspecting notes before and after the main AI speaks.
1. The sneaky note trick
Attackers try to fool AI models by hiding secret instructions inside normal requests.
Why does this trick work?
Large models read everything as one long conversation. If an attacker tells the model to forget its rules, an unguarded model might obey.
2. The guard at the entrance
A small firewall model checks the question first and throws out tricks before the main brain sees them.
What does the guard check?
The guard model is trained on attack patterns. It scans for jailbreak attempts, toxic content, and attempts to override system safeguards.
3. The guard at the exit
Before sending the answer back, the exit guard makes sure no private passwords or secrets slipped out.
Why inspect the answer too?
Even if an attack bypassed the entrance guard, the exit guard prevents secret API keys or customer records from leaving the system.
Longer version: Reference Entry on Firewall LLM
Sources: Inan et al., Llama Guard (2023); OWASP Top 10 for LLMs.