You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I maintain the MIT project. Learners trace an untrusted memory write through later recall, summary laundering, repetition, retrieval flooding and preference grafting, then compare origin-aware defense and bounded repair. There is a dependency-free reference store, optional ChromaDB/sqlite-vec adapters, tests and an offline browser replay.
It adds a persistent-memory boundary exercise rather than another jailbreak prompt collection. The targets are synthetic and the agent is scripted and deterministic, not an LLM. The lab teaches mechanics; it does not establish real-model safety or claim a production defense. Its preference-grafting case explicitly exposes the limit of repair when there is no owner policy for that subject.
I read the contribution criteria and checked the current list and issue/PR suggestions for duplicates. Since the current categories do not have a tools/labs section, I am asking about fit before proposing a new category or placing it under courses. If it fits, which existing section would you prefer?
Activity