Poisoning
How untrusted input can become persistent memory, re-enter a later decision, and survive naive deletion through derived state.
Links containing ?t= open the video at a specific second.
The ideas to retain
Storing data does not make it trustworthy
Storing an output in a database does not make it trustworthy. An agent's memory should record its origin, date, scope, permissions and expiration policy. Without those properties, the…
Persistent memory is already a measurable attack surface
Evidence published in 2026 lets us analyze this surface much more directly than the earliest work on backdoors in model weights.
Dangerous behavior can also remain hidden in the model
The Sleeper Agents study built test models that wrote safe code when the prompt indicated 2023 and vulnerable code when it indicated 2024. The demonstration does not describe a commercial…
What this video covers
This video does not yet have a reviewed synchronized transcript. This editorial context describes its content without presenting it as spoken audio or literal on-screen text.
How untrusted input can become persistent memory, re-enter a later decision, and survive naive deletion through derived state.
- Storing data does not make it trustworthy: Storing an output in a database does not make it trustworthy. An agent's memory should record its origin, date, scope, permissions and expiration policy. Without those properties, the…
- Persistent memory is already a measurable attack surface: Evidence published in 2026 lets us analyze this surface much more directly than the earliest work on backdoors in model weights.
- Dangerous behavior can also remain hidden in the model: The Sleeper Agents study built test models that wrote safe code when the prompt indicated 2023 and vulnerable code when it indicated 2024. The demonstration does not describe a commercial…
