Why usual testing misses this
Application and model testing rarely touch the store itself. Retrieval can surface poisoned documents or another tenant's data even when the model and prompts look safe.
Vector store red teaming
Find out what your vector store gives up when someone is trying.
VecShield runs authorised adversarial testing against the vector stores and retrieval layer behind your RAG application: the same collections, the same namespaces, the same access paths your application uses. You get reproducible cases with the exact inputs, what was retrieved, and what has to change.
How an engagement runs
- 1
Scope
We agree the system, the corpus, the tools in reach and the rules of engagement, in writing.
- 2
Test
Authorised adversarial runs against the real stores: document poisoning and injection, sensitive-data recovery through retrieval, cross-tenant and cross-namespace leakage.
- 3
Reproduce
Every finding ships with the exact inputs that worked and the response that came back.
- 4
Remediate
What has to change in the store, the ingestion path or access controls, and a re-test of the fixes.
What we test and what you receive
Attack classes
Document poisoning and injection, sensitive-data recovery through retrieval, and cross-tenant or cross-namespace leakage, run against the collections your application actually uses.
Reproducible cases
Each finding includes the exact inputs, what was retrieved, and the response that came back, so your team can replay it and confirm the fix.
Authorisation and limits
Testing runs only within written, agreed scope and rules of engagement. We report what was tested and what was not, and do not claim coverage we did not achieve.
Scope an engagement.
Tell us what the vector stores and retrieval layer expose and we'll tell you what we'd try.