Agent memory poisoning
Named by 3 kept items in this workspace.
Search every workspace for Agent memory poisoning. It matches the name as written, so it can find other things that share it.
The items that name it 3
Injecting false or malicious content into episodic, semantic, or procedural agent memory; plain false statements (no adversarial crafting) poisoning 1.2% of a corpus dropped accuracy from 85% to 30%; screening caught 0 of 360 poisoned memories despite catching 83% of indirect prompt injection
MINJA / MemoryGraft
A malicious instruction crafted by an attacker is stored in an agent's long-term memory, recalled much later, and quietly shapes a real action
All techniques in this workspace.