WeaknessASC-2026-0006Major weakness
Memory and context poisoning
HighWorkarounds only
Attackers write false facts or instructions into long-term memory or RAG stores that later sessions trust.
How to detect it
- Persistent memory or a vector store accepts writes with no validation
- Retrieved documents contain instruction phrases or hidden text
Workarounds
- Validate and attribute every memory write (who, when, from which source)
- Expire memory entries and review high-impact ones
- Isolate stores per tenant and per trust level
- Retrieve with provenance so untrusted chunks are labeled as data
No general fix exists for poisoned retrieval; controls limit how far a poisoned entry can spread.
How to investigate
- Find memory or RAG entries written from untrusted sources around the incident
- Run the content inspector on suspicious entries
- List sessions that retrieved the entry, purge it and re-index