CVE-2026-25211: Llama Stack security flaw enables

CISO Take

Llama Stack exposed pgvector database credentials in plaintext initialization logs, affecting any deployment using pgvector as a vector store backend. Patch to llama-stack >= 0.4.4 immediately and rotate all pgvector passwords — assume any credentials logged prior to patching are compromised. Audit log access controls: if logs reached a SIEM, cloud log aggregator, or shared storage, treat the pgvector database as fully exposed.

What is the risk?

Inherent CVSS is low (3.2) with local attack vector and high complexity, but the practical risk is elevated in typical enterprise AI deployments where logs are centralized in Splunk, Elastic, or cloud log services accessible to a broader audience than the application host. The scope change (S:C) in the vector indicates the credential exposure crosses into the pgvector database security boundary. EPSS near zero reflects no current weaponization, but credential-in-log vulnerabilities are trivial to exploit once an attacker gains any log read access — which is a common misconfiguration in AI dev environments.

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
Llama Stack	pip	< 0.4.4	`0.4.4`
8.4K 14 dependents Pushed 3d ago 100% patched ~0d to patch Full package profile →

Do you use Llama Stack? You're affected.

How severe is it?

CVSS 3.1

3.2 / 10

EPSS

0.2%

chance of exploitation in 30 days

Higher than 12% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

No known exploitation

Sophistication

Trivial

What is the attack surface?

AV Local

AC High

PR None

UI None

S Changed

C Low

I None

A None

What should I do?

1 step

1) PATCH: Upgrade llama-stack to >= 0.4.4 immediately. 2) ROTATE: Change pgvector passwords on all affected instances regardless of perceived log exposure. 3) AUDIT LOGS: Search existing log archives for 'pgvector', 'password', 'postgres://', or similar connection string patterns — check SIEM, CloudWatch, Elastic, Splunk. 4) RESTRICT: Apply least-privilege access to application logs; logs containing initialization output should not be readable by application users or broad ops teams. 5) DETECT: Add a log monitoring rule for pgvector/PostgreSQL connection strings appearing in application logs. 6) VERIFY: Confirm no unauthorized connections to the pgvector database in the period between initial deployment and patching by reviewing PostgreSQL pg_stat_activity history or audit logs.

What does CISA's SSVC say?

Decision Track

Exploitation none

Automatable No

Technical Impact partial

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Data Leakage Data Extraction RAG Framework AML.T0037 - Data from Local System AML.T0055 - Unsecured Credentials AML.T0070 - RAG Poisoning AML.T0085.000 - RAG Databases

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, Robustness and Cybersecurity

ISO 42001

A.6.1.2 - Information Security in AI System Lifecycle A.9.4 - Logging and monitoring of AI systems

NIST AI RMF

GOVERN 1.6 - Policies and practices address AI risks across the lifecycle MANAGE 2.2 - Mechanisms are in place to deal with AI risk and to recover from incidents MANAGE-2.2 - Risk Response — Treatment of Identified AI Risks

OWASP LLM Top 10

LLM02 - Sensitive Information Disclosure LLM06 - Sensitive Information Disclosure LLM08 - Vector and Embedding Weaknesses

Frequently Asked Questions

What is CVE-2026-25211?

Llama Stack exposed pgvector database credentials in plaintext initialization logs, affecting any deployment using pgvector as a vector store backend. Patch to llama-stack >= 0.4.4 immediately and rotate all pgvector passwords — assume any credentials logged prior to patching are compromised. Audit log access controls: if logs reached a SIEM, cloud log aggregator, or shared storage, treat the pgvector database as fully exposed.

Is CVE-2026-25211 actively exploited?

No confirmed active exploitation of CVE-2026-25211 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-25211?

1) PATCH: Upgrade llama-stack to >= 0.4.4 immediately. 2) ROTATE: Change pgvector passwords on all affected instances regardless of perceived log exposure. 3) AUDIT LOGS: Search existing log archives for 'pgvector', 'password', 'postgres://', or similar connection string patterns — check SIEM, CloudWatch, Elastic, Splunk. 4) RESTRICT: Apply least-privilege access to application logs; logs containing initialization output should not be readable by application users or broad ops teams. 5) DETECT: Add a log monitoring rule for pgvector/PostgreSQL connection strings appearing in application logs. 6) VERIFY: Confirm no unauthorized connections to the pgvector database in the period between initial deployment and patching by reviewing PostgreSQL pg_stat_activity history or audit logs.

What systems are affected by CVE-2026-25211?

This vulnerability affects the following AI/ML architecture patterns: RAG pipelines, vector databases, LLM application frameworks, AI agent frameworks.

What is the CVSS score for CVE-2026-25211?

CVE-2026-25211 has a CVSS v3.1 base score of 3.2 (LOW). The EPSS exploitation probability is 0.22%.

What is the AI security impact?

Affected AI Architectures

RAG pipelinesvector databasesLLM application frameworksAI agent frameworks

MITRE ATLAS Techniques

AML.T0037 Data from Local System

AML.T0055 Unsecured Credentials

AML.T0070 RAG Poisoning

AML.T0085.000 RAG Databases

Compliance Controls Affected

EU AI Act: Article 15

ISO 42001: A.6.1.2, A.9.4

NIST AI RMF: GOVERN 1.6, MANAGE 2.2, MANAGE-2.2

OWASP LLM Top 10: LLM02, LLM06, LLM08

What are the technical details?

Original Advisory

Llama Stack (aka llama-stack) before 0.4.0rc3 does not censor the pgvector password in the initialization log.

Exploitation Scenario

An attacker with read access to Llama Stack application logs — via a compromised CI/CD pipeline, misconfigured S3 bucket storing logs, over-permissioned CloudWatch log group, or insider access — extracts the pgvector connection string from the initialization log entry. The credential is valid for direct TCP access to the PostgreSQL/pgvector instance. The attacker connects directly to the vector database, bypassing Llama Stack entirely, and issues SQL queries against the vector tables to exfiltrate the entire embedding store and associated metadata (document chunks, source references, user query data if stored). In a second-stage attack, the attacker inserts crafted embeddings that poison RAG retrieval, causing the LLM to return attacker-controlled content to end users without any visible indicators of compromise.

Weaknesses (CWE)

CWE-532 Insertion of Sensitive Information into Log File Primary

CWE-532 — Insertion of Sensitive Information into Log File: The product writes sensitive information to a log file.

[Architecture and Design, Implementation] Consider seriously the sensitivity of the information written into log files. Do not write secrets into the log files.
[Distribution] Remove debug log files before deploying the application into production.

Source: MITRE CWE corpus.