GASLIT

The problem

You can attack an AI agent by slowly feeding it false memories. Each message looks harmless on its own, so normal guardrails let it through, and the pattern only shows up across the agent's memory over time.

What I built

I built a layer that sits between the agent's memory database and the model and checks every memory before the agent uses it. Four agents cross check new entries against the rest, verify nothing has been tampered with, and quarantine anything suspicious.

What it does

• Checks at the Last Moment. It inspects each memory right when the agent retrieves it, after the attack has already slipped past guardrails that only look at single messages.
• Four Agents on Guard. The agents never call each other directly. They coordinate entirely through the same MongoDB database they are protecting.
• Slow Poison Shows Up as Drift. It fingerprints the queries that touch each memory, so an attack spread across weeks still stands out statistically.
• A Human Decides. The system never acts on its own. Every flag becomes a quarantine report for a person to review.

Gallery

The full engineering is on GitHub.