Browse / RAG and Vector Store Weaknesses
RAG and Vector Store Weaknesses
Flaws in retrieval pipelines, embeddings and vector databases. 36 records. Reference: OWASP LLM08 · ATLAS AML.T0070.
- Authentication and access control 10
- SQL and query injection 8
- Code and command execution 3
- Other weaknesses 3
- Server-side request forgery 3
- Path traversal and file access 2
- Denial of service 2
- Information exposure 2
- Cross-site request forgery 1
- RAG Poisoning 1
- Gather RAG-Indexed Targets 1
- mem0ai Databricks vector store vulnerable to SQL injection Stored memory text or metadata reaches unescaped string interpolation in the Databricks vector store's insert and update statements, so attacker-chosen SQL runs on the warehouse.
- PraisonAI: PGVector and Cassandra knowledge stores interpolate vector dimensions into DDL
- Open WebUI: Inaccessible knowledge bases are exposed through the built-in knowledge tool on most vector backends
- Open WebUI: Cross-user file content disclosure via request-scoped direct model knowledge metadata Knowledge attached to an inline model skips the access check, so an authenticated user who knows another user's file ID can retrieve its indexed content.
- Open WebUI: Client-side SSRF via unrestricted external resource loading in Vega/Vega-Lite chart rendering
- Open WebUI: SSRF into internal services via unvalidated sub-resource requests in the Playwright web loader
- Flowise RCE via SQLite Record Manager Node
- Open WebUI: Model meta.knowledge read-only file access can be upgraded to file write/delete A model's knowledge-file metadata lets a read-only grant satisfy later write and delete checks, because an earlier fix validated only read access.
- Open WebUI: ReDoS in skill-mention regexes causes whole-instance DoS on default config
- Open WebUI: `WEB_FETCH_FILTER_LIST` host allow/block filter bypassable via URL path and non-label-boundary matching
- Langflow: BaseFileComponent-based nodes arbitrary file read with RCE exploit A user-supplied file ingested by the Read File node can direct it to read arbitrary absolute paths, exposing secrets that enable token forgery and code execution.
- Open WebUI: RAG ACL Bypass in Milvus Multitenancy Mode
- Open WebUI: Cross-user file disclosure via /api/chat/completions image_url field
- ChromaDB allows any authenticated users to arbitrarily read, write, update, or delete data in any tenant's collection The vector database omits authorisation checks, so any authenticated user can read and modify collections belonging to other tenants.
- agno contains a SQL injection vulnerability Metadata keys and values passed to the ClickHouse vector store delete method are interpolated into SQL, letting attackers alter queries and read data.
- Open WebUI: Cross-User File Access via Unchecked file_id in Folder Knowledge and Knowledge-Base Attach Endpoints
- FlowiseAI: Vector Store No Permission Checks
- PraisonAI knowledge-store backends interpolate unvalidated collection names into SQL and CQL queries Collection names passed to SQL and CQL knowledge-store backends are concatenated into table and index identifiers without validation, enabling query injection.
- Open WebUI vulnerable to Global Knowledge Base Enumeration via knowledge-bases Meta-Collection The collection access check only covers certain name prefixes, so querying the system knowledge-base collection lists every knowledge base's metadata.
- Open WebUI has Knowledge Base Destruction and RAG Poisoning via Unauthorized Collection Overwrite A web-retrieval endpoint accepts any collection name and overwrites it without ownership checks, so other users' vector collections can be deleted or poisoned.
- Open WebUI has Unauthorized File and Knowledge Base Content Access via RAG Vector Search
- Open WebUI's responses passthrough endpoint lacks access control authorization
- Spring AI has a VectorStore FilterExpression Converter injection
- RAGAS has SSRF via Multi-Modal Faithfulness Collections Module The evaluator fetches URLs found in retrieved contexts without the SSRF guard added elsewhere, so attacker-supplied context can make the server send requests.
- Flowise: Path Traversal in Vector Store basePath
- Spring AI: SpEL injection is triggered when a user-supplied value is used as a filter expression key User-supplied values used as keys in SimpleVectorStore filter expressions are evaluated as SpEL, allowing arbitrary code execution.
- JSONPath Injection in Spring AI Vector Stores FilterExpressionConverter
- Vanna has a SQL injection in the remove_training_data function The ID argument of a BigQuery vector store delete function is concatenated into SQL, letting callers inject query clauses.
- Lobe Chat has IDOR in Knowledge Base File Removal that Allows Cross User File Deletion
- LlamaIndex vulnerable to data loss through hash collisions in its DocugamiReader class Chunk IDs derived from MD5 of chunk text collide for structurally different chunks, so one overwrites the other in the index.
- llama_index vulnerable to SQL Injection
- LlamaIndex Retrievers Integration: DuckDBRetriever SQL Injection Untrusted query text is concatenated into DuckDB SQL, letting an attacker load the shellfs extension and run OS commands.
- Open WebUI Unauthenticated Multipart Boundary Denial of Service (DoS) Vulnerability
- Open WebUI Vulnerable to Cross-Site Request Forgery (CSRF)
- Financial Transaction Hijacking with M365 Copilot as an Insider
- Morris II Worm: RAG-Based Attack