CVE-2026-104873: LangGraph SDK: auth decorator flaw enables cross-user access

GHSA-fvww-7h3r-vfhpHIGH
Published October 2, 2026
CISO Take

The langgraph-sdk Python package (0.1.45 through 0.4.3) has an authorization bug. The @auth.on.threads, @auth.on.assistants and @auth.on.crons decorators silently ignore the actions argument and register the handler for every action on the resource. Because that wildcard handler is selected before broader fallback handlers, an authenticated user can skip ownership or permission checks and read, modify or delete another user's threads, assistants or cron jobs. Exploitation likelihood is low to moderate (EPSS 0.25%, top 85th percentile), and there is no public exploit, Nuclei template or KEV listing. However, the package has about 4,000 downstream dependents, and the exposed objects often hold conversation history, agent configuration and scheduled automations. Upgrade langgraph-sdk to 0.4.4 or later. Until then, audit any handler registered with actions= and make sure it enforces ownership checks for every action it can receive.

Sources:NVDGitHub AdvisoryEPSSATLASlangchain-ai/langgraph security advisory

What is the risk?

Severity is high, although no CVSS vector is published. The attacker needs a valid authenticated account on the LangGraph API server, which limits exposure to multi-tenant or multi-user deployments. Exploitation is low-complexity once authenticated, because the flaw is a logic error in handler registration and needs no special payload. Exploitation likelihood is modest (EPSS 0.00253, no known exploits, not in KEV). Confidentiality and integrity impact on tenant data can be significant. Only Python deployments that use the actions argument on the affected decorators are vulnerable, and a deployment is not exposed if its selected handler independently enforces all required checks for every action.

How does the attack unfold?

Initial Access
Attacker obtains a legitimate authenticated account on a LangGraph API server that uses resource-scoped auth handlers.
AML.T0012
Authorization Bypass
The action-scoped handler is registered for every action, so ownership and fallback permission checks are skipped.
Collection
Attacker reads other users' threads, assistants and cron definitions by ID.
AML.T0036
Impact
Attacker modifies or deletes other tenants' resources, causing data exposure, tampered agent behavior and service disruption.

What systems are affected?

PackageEcosystemVulnerable RangePatched
LangGraphpip—No patch
45.0K4.2K dependentsPushed 3d ago77% patched~9d to patchFull package profile →
LangGraphpip>= 0.1.45, <= 0.4.30.4.4
45.0K4.2K dependentsPushed 3d ago77% patched~9d to patchFull package profile →

How severe is it?

CVSS 3.1
N/A
EPSS
0.3%
chance of exploitation in 30 days
Higher than 17% of all CVEs
Exploitation Status
No known exploitation
Sophistication
Trivial

What should I do?

5 steps
  1. Upgrade langgraph-sdk to 0.4.4 or later (fix commit 5a77be5e in langchain-ai/langgraph) and redeploy all LangGraph API servers.

  2. Inventory auth handlers in code and flag any use of @auth.on.threads, @auth.on.assistants or @auth.on.crons with an actions= argument.

  3. Until patched, rework those handlers to enforce ownership and permission checks for every action they receive, or fall back to generic resource handlers without actions filtering.

  4. Review API access logs for authenticated users reading, updating or deleting thread, assistant or cron IDs that do not belong to them, and for unusual deletions.

  5. Add integration tests that assert cross-user access is denied for each action (read, create, update, delete, search).

What does CISA's SSVC say?

DecisionTrack
Exploitationnone
AutomatableNo
Technical Impactpartial

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:

EU AI Act
Art. 15-Accuracy, robustness and cybersecurity
ISO 42001
A.6.2.6-AI system operation and monitoring
NIST AI RMF
MANAGE 2.2-Mechanisms to sustain the value of deployed AI systems
OWASP LLM Top 10
LLM02-Sensitive Information Disclosure

Frequently Asked Questions

What is CVE-2026-104873?

The langgraph-sdk Python package (0.1.45 through 0.4.3) has an authorization bug. The @auth.on.threads, @auth.on.assistants and @auth.on.crons decorators silently ignore the actions argument and register the handler for every action on the resource. Because that wildcard handler is selected before broader fallback handlers, an authenticated user can skip ownership or permission checks and read, modify or delete another user's threads, assistants or cron jobs. Exploitation likelihood is low to moderate (EPSS 0.25%, top 85th percentile), and there is no public exploit, Nuclei template or KEV listing. However, the package has about 4,000 downstream dependents, and the exposed objects often hold conversation history, agent configuration and scheduled automations. Upgrade langgraph-sdk to 0.4.4 or later. Until then, audit any handler registered with actions= and make sure it enforces ownership checks for every action it can receive.

Is CVE-2026-104873 actively exploited?

No confirmed active exploitation of CVE-2026-104873 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-104873?

1. Upgrade langgraph-sdk to 0.4.4 or later (fix commit 5a77be5e in langchain-ai/langgraph) and redeploy all LangGraph API servers. 2. Inventory auth handlers in code and flag any use of @auth.on.threads, @auth.on.assistants or @auth.on.crons with an actions= argument. 3. Until patched, rework those handlers to enforce ownership and permission checks for every action they receive, or fall back to generic resource handlers without actions filtering. 4. Review API access logs for authenticated users reading, updating or deleting thread, assistant or cron IDs that do not belong to them, and for unusual deletions. 5. Add integration tests that assert cross-user access is denied for each action (read, create, update, delete, search).

What systems are affected by CVE-2026-104873?

This vulnerability affects the following AI/ML architecture patterns: agent frameworks, multi-tenant LLM application backends, LangGraph API servers, stateful agent workflows, scheduled agent automation.

What is the CVSS score for CVE-2026-104873?

No CVSS score has been assigned yet.

What is the AI security impact?

Affected AI Architectures

agent frameworksmulti-tenant LLM application backendsLangGraph API serversstateful agent workflowsscheduled agent automation

MITRE ATLAS Techniques

AML.T0012Valid Accounts
AML.T0036Data from Information Repositories

Compliance Controls Affected

EU AI Act: Art. 15
ISO 42001: A.6.2.6
NIST AI RMF: MANAGE 2.2
OWASP LLM Top 10: LLM02

What are the technical details?

Original Advisory

LangGraph Python SDK is used to connect to running LangGraph API servers, manage assistants, threads and stream runs from Python applications. From 0.1.45 until 0.4.4, the langgraph-sdk resource-scoped authorization decorators @auth.on.threads, @auth.on.assistants, and @auth.on.crons ignore the actions argument and register the selected handler for every action on the resource. Because that wildcard resource handler is selected before broader fallback handlers, an authenticated user may bypass fallback action, ownership, or permission checks and read, update, or delete another user's resource. Only Python deployments using actions on the affected decorators are vulnerable, and a deployment remains protected when the selected handler independently enforces all required checks for every action it receives. This issue is fixed in version 0.4.4.

Exploitation Scenario

In a multi-tenant SaaS offering built on LangGraph, an attacker signs up for a legitimate low-privilege account. The operator restricted deletion and cross-user reads with action-specific handlers, such as @auth.on.threads with actions=["read"], and relies on a broader fallback handler for the rest. Because the actions argument is ignored, the narrow handler runs for every action and skips the ownership check. The attacker enumerates or guesses thread IDs and reads other customers' conversations, which may contain proprietary data or credentials pasted into prompts. They then modify or delete other tenants' assistants and cron jobs to disrupt service or tamper with agent behavior.

Weaknesses (CWE)

CWE-863 — Incorrect Authorization: The product performs an authorization check when an actor attempts to access a resource or perform an action, but it does not correctly perform the check.

  • [Architecture and Design] Divide the product into anonymous, normal, privileged, and administrative areas. Reduce the attack surface by carefully mapping roles with data and functionality. Use role-based access control (RBAC) [REF-229] to enforce the roles at the appropriate boundaries. Note that this approach may not protect against horizontal authorization, i.e., it will not protect a user from attacking others with the same role.
  • [Architecture and Design] Ensure that access control checks are performed related to the business logic. These checks may be different than the access control checks that are applied to more generic resources such as files, connections, processes, memory, and database records. For example, a database may restrict access for medical records to a specific database user, but each record might only be intended to be accessible to the patient and the patient's doctor [REF-7].

Source: MITRE CWE corpus.

Timeline

Published
October 2, 2026
Last Modified
October 5, 2026
First Seen
October 2, 2026

Related Vulnerabilities