agmiAgent Memory Integrity GitHub

Front-door attack

Planted fact

Write a false fact through the tool's own API and see whether recall serves it as the user's own.

What it looks like in the world: The cheapest attack there is: talk to the agent and wait.

attack id memory_injection · version 3 · attacker: write-access

Every store, this edit

surfaced 7

Reproduce

pip install agent-memory-integrity
python agmi/full_runner.py --json results/scorecard.json   # every row; the memory_injection column is this page

All 15 edits and attacks · The scorecard