CVE-2026-103484: pgvector: IVFFlat build OOB write enables code execution
HIGHpgvector versions before 0.8.7 contain an out-of-bounds write in the IVFFlat index build routine, which lets a low-privileged database user corrupt memory and potentially run arbitrary code inside the PostgreSQL process. The CVSS score is 8.8 (high), and about 400 downstream packages depend on pgvector, so many RAG and semantic-search backends inherit the exposure. Near-term exploitation risk looks modest: EPSS is 0.00422, the CVE is not in CISA KEV, no public exploit or Nuclei template exists, and CISA SSVC says TRACK. The weak point is the privilege requirement, because an attacker needs a database account that can build indexes, which usually means stolen application credentials or SQL injection in the AI application. Upgrade pgvector to 0.8.7 or later, restrict CREATE INDEX rights on vector columns to trusted administrators, and treat this as a patch-in-cycle item rather than an emergency, since this package already has 3 other CVEs.
What is the risk?
Overall risk is MEDIUM-HIGH for organizations running pgvector. Severity is high (CVSS 8.8), but practical exploitability is tempered by the low-privilege authentication requirement, the lack of a public exploit, EPSS of 0.42%, no KEV listing and an SSVC decision of TRACK. Impact is serious if exploited: memory corruption in the PostgreSQL backend can give code execution as the database service account, with full confidentiality, integrity and availability loss for the database. Exposure is highest where application tiers, notebooks or multi-tenant services hand database credentials to many users or where SQL injection could reach DDL statements. Confidence: HIGH on the vulnerability facts (NVD description, vendor commit), MEDIUM on real-world exploitability, because exploit details are not public.
How does the attack unfold?
What systems are affected?
| Package | Ecosystem | Vulnerable Range | Patched |
|---|---|---|---|
| pgvector | pip | — | No patch |
Do you use pgvector? You're affected.
How severe is it?
What is the attack surface?
What should I do?
1 step1) Upgrade the pgvector extension to 0.8.7 or later (the fix is in upstream commit ee00d39 / issue #1036) and run ALTER EXTENSION vector UPDATE in each database that uses it. 2) Until patched, limit the privilege to create or rebuild IVFFlat indexes (CREATE INDEX, REINDEX) to a small set of trusted administrative roles, and avoid IVFFlat builds from application or notebook accounts. 3) Audit database roles used by RAG services, data science notebooks and ETL jobs, and enforce least privilege with separate credentials for read-only query workloads. 4) Detection: log DDL statements (log_statement = ddl) and alert on unexpected CREATE INDEX ... USING ivfflat from non-admin roles, and watch for backend crashes (signal 11 / server process terminated abnormally) in PostgreSQL logs. 5) Run PostgreSQL under a non-root service account with OS-level hardening (SELinux/AppArmor, no unnecessary outbound network access) to contain post-exploitation. 6) Inventory pgvector versions across environments, including managed database offerings, and confirm the provider has shipped the fixed build.
What does CISA's SSVC say?
Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.
How is it classified?
Which compliance frameworks are affected?
This CVE is relevant to:
Frequently Asked Questions
What is CVE-2026-103484?
pgvector versions before 0.8.7 contain an out-of-bounds write in the IVFFlat index build routine, which lets a low-privileged database user corrupt memory and potentially run arbitrary code inside the PostgreSQL process. The CVSS score is 8.8 (high), and about 400 downstream packages depend on pgvector, so many RAG and semantic-search backends inherit the exposure. Near-term exploitation risk looks modest: EPSS is 0.00422, the CVE is not in CISA KEV, no public exploit or Nuclei template exists, and CISA SSVC says TRACK. The weak point is the privilege requirement, because an attacker needs a database account that can build indexes, which usually means stolen application credentials or SQL injection in the AI application. Upgrade pgvector to 0.8.7 or later, restrict CREATE INDEX rights on vector columns to trusted administrators, and treat this as a patch-in-cycle item rather than an emergency, since this package already has 3 other CVEs.
Is CVE-2026-103484 actively exploited?
No confirmed active exploitation of CVE-2026-103484 has been reported, but organizations should still patch proactively.
How to fix CVE-2026-103484?
1) Upgrade the pgvector extension to 0.8.7 or later (the fix is in upstream commit ee00d39 / issue #1036) and run ALTER EXTENSION vector UPDATE in each database that uses it. 2) Until patched, limit the privilege to create or rebuild IVFFlat indexes (CREATE INDEX, REINDEX) to a small set of trusted administrative roles, and avoid IVFFlat builds from application or notebook accounts. 3) Audit database roles used by RAG services, data science notebooks and ETL jobs, and enforce least privilege with separate credentials for read-only query workloads. 4) Detection: log DDL statements (log_statement = ddl) and alert on unexpected CREATE INDEX ... USING ivfflat from non-admin roles, and watch for backend crashes (signal 11 / server process terminated abnormally) in PostgreSQL logs. 5) Run PostgreSQL under a non-root service account with OS-level hardening (SELinux/AppArmor, no unnecessary outbound network access) to contain post-exploitation. 6) Inventory pgvector versions across environments, including managed database offerings, and confirm the provider has shipped the fixed build.
What systems are affected by CVE-2026-103484?
This vulnerability affects the following AI/ML architecture patterns: vector databases, RAG pipelines, semantic search, agent memory stores, embedding storage on PostgreSQL.
What is the CVSS score for CVE-2026-103484?
CVE-2026-103484 has a CVSS v3.1 base score of 8.8 (HIGH). The EPSS exploitation probability is 0.47%.
What is the AI security impact?
Affected AI Architectures
MITRE ATLAS Techniques
AML.T0010.001AI SoftwareAML.T0049Exploit Public-Facing ApplicationCompliance Controls Affected
What are the technical details?
Original Advisory
IVFFlat index build in pgvector before 0.8.7 allows a database user to write data out-of-bounds, which can lead to arbitrary code execution.
Exploitation Scenario
An attacker obtains low-privilege credentials for a database used by an internal RAG application, for example through a leaked connection string in a notebook repository or SQL injection in a document-search endpoint. With a role that is allowed to create indexes on a vector table, they insert crafted vector data and trigger an IVFFlat index build, causing an out-of-bounds write in the PostgreSQL backend. They turn the memory corruption into code execution as the postgres user, then read all embeddings and source documents, harvest credentials stored in the database, and pivot to the application network. They may also alter stored embeddings to manipulate what the LLM retrieves for later users.
Weaknesses (CWE)
CWE-1284Improper Validation of Specified Quantity in InputCWE-1284Improper Validation of Specified Quantity in InputCWE-787Out-of-bounds WriteCWE-787Out-of-bounds WriteCWE-1284 — Improper Validation of Specified Quantity in Input: The product receives input that is expected to specify a quantity (such as size or length), but it does not validate or incorrectly validates that the quantity has the required properties.
- [Implementation] Assume all input is malicious. Use an "accept known good" input validation strategy, i.e., use a list of acceptable inputs that strictly conform to specifications. Reject any input that does not strictly conform to specifications, or transform it into something that does. When performing input validation, consider all potentially relevant properties, including length, type of input, the full range of acceptable values, missing or extra inputs, syntax, consistency across related fields, and conformance to business rules. As an example of business rule logic, "boat" may be syntactically valid because it only contains alphanumeric characters, but it is not valid if the input is only expected to contain colors such as "red" or "blue." Do not rely exclusively on looking for malicious or malformed inputs. This is likely to miss at least one undesirable input, especially if the code's environment changes. This can give attackers enough room to bypass the intended validation. However, denylis
Source: MITRE CWE corpus.
CVSS Vector
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:HReferences
Timeline
Related Vulnerabilities
CVE-2026-180228.8pgvector: integer wraparound in index build enables RCE
Same package: pgvectorCVE-2026-554057.6langchain4j: SQL injection in vector store filters
Same package: pgvectorCVE-2026-1253265.8pgEdge RAG Server: API key fragment leak via error text
Same package: pgvectorCVE-2026-48148Budibase: SSRF via VectorDB host exposes cloud metadata
Same package: pgvectorCVE-2026-2185810.0n8n: Input Validation flaw enables exploitation
Same attack type: Code Execution