assessing-vector-and-embedding-weaknesses - Assess security weaknesses in vector stores and embeddings
Test vector stores for embedding inversion, cross-tenant leakage, membership inference, and knowledge-base poisoning.
Tags
Updated: 2026-09-24Capabilities
Typical Inputs
Typical Outputs
What this skill does
- Measure embedding inversion exposure
- Run membership inference probes
- Test multi-tenant isolation controls
- Inject test chunks for poisoning
- Detect indirect prompt injection content
Inputs
- Target embedding endpoint
- Vector store test collection
- Test queries and text samples
- Scope authorization documentation
Outputs
- Embedding inversion similarity scores
- Membership inference score deltas
- Multi-tenant isolation leak reports
- Poisoning retrieval dominance results
- Indirect prompt injection detection flags
Requirements
- Authorization for embedding endpoint and vector store
- Python 3.10 or higher
- Read and write access to a test collection