Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-111767 Overview
CVE-2026-111767 affects the Model Context Protocol (MCP) TypeScript SDK, the official SDK for building MCP servers and clients. The experimental tasks feature exposed by taskStore failed to validate which session created a task. Clients sharing a single InMemoryTaskStore instance could list, read, and cancel tasks belonging to other sessions. The flaw is classified under [CWE-639] Authorization Bypass Through User-Controlled Key. It affects SDK versions 1.24.0 through versions prior to 1.32.0 and is resolved in 1.32.0.
Critical Impact
Any MCP client sharing an InMemoryTaskStore with other sessions can enumerate, read task output, and cancel tasks owned by other sessions, breaking tenant isolation in multi-session MCP server deployments.
Affected Products
- MCP TypeScript SDK versions 1.24.0 through 1.31.x
- MCP servers using the experimental taskStore feature with a shared InMemoryTaskStore
- Multi-tenant MCP deployments relying on session-based isolation
Discovery Timeline
- 2026-10-20 - CVE-2026-111767 published to NVD
- 2026-10-20 - Last updated in NVD database
Technical Details for CVE-2026-111767
Vulnerability Analysis
The MCP TypeScript SDK introduced an experimental tasks subsystem in version 1.24.0 to support long-running task execution across MCP sessions. The subsystem exposes a TaskStore interface with methods such as createTask, listTasks, getTask, and cancelTask. The reference implementation, InMemoryTaskStore, holds all tasks in a shared process-wide map.
The TaskStore methods did not accept or enforce a sessionId parameter. Any session connected to a server that shares one InMemoryTaskStore instance could invoke these methods and operate on tasks created by other sessions. This breaks the implicit tenant boundary that MCP server operators expect when multiplexing clients over a single process.
Exploitation requires only a connected client with permission to invoke the experimental tasks API. An attacker-controlled session can enumerate other sessions' task identifiers via listTasks, read task state and output via getTask, and disrupt workloads via cancelTask.
Root Cause
The root cause is a missing authorization check keyed on session identity. The TaskStore contract did not require implementations to associate tasks with the creating session or filter results by requester. The in-memory implementation therefore returned all stored tasks regardless of caller.
Attack Vector
An attacker needs network access to an MCP server that uses the experimental tasks feature with a shared InMemoryTaskStore across sessions. No authentication beyond an active MCP session is required. The attacker calls listTasks to discover task IDs created by other sessions and then issues getTask or cancelTask against those IDs.
* Similar to Transport, this allows pluggable task storage implementations
* (in-memory, database, distributed cache, etc.).
*
+ * Every method receives the sessionId of the request it serves. A store shared by several
+ * sessions should record the sessionId passed to createTask and treat a task as not found
+ * (and leave it out of listTasks) when a different sessionId asks for it.
+ *
* @experimental
*/
export interface TaskStore {
Source: GitHub Commit 727075b. The patch updates the TaskStore interface so every method receives the request's sessionId, and requires shared stores to scope task visibility by that identifier.
Detection Methods for CVE-2026-111767
Indicators of Compromise
- Unexpected tasks/list, tasks/get, or tasks/cancel MCP method calls originating from sessions that did not create the referenced tasks
- Tasks transitioning to a cancelled state without a corresponding cancel request from their creating session
- Access patterns where a single session reads task output IDs it never created
Detection Strategies
- Inventory MCP server deployments and identify any instance running @modelcontextprotocol/sdk versions 1.24.0 through 1.31.x with the experimental tasks feature enabled
- Review MCP server logs for cross-session task access by correlating sessionId on inbound requests with the sessionId recorded at task creation
- Add application-level logging around TaskStore method invocations to capture caller session and target task identifiers
Monitoring Recommendations
- Alert on anomalous volumes of listTasks calls from a single session, which may indicate task enumeration
- Monitor task cancellation rates and investigate cancellations not initiated by the task owner
- Track dependency versions of @modelcontextprotocol/sdk in CI and alert when any build pins a version below 1.32.0
How to Mitigate CVE-2026-111767
Immediate Actions Required
- Upgrade @modelcontextprotocol/sdk to version 1.32.0 or later in all MCP server and client projects
- Audit custom TaskStore implementations to ensure every method filters by the request sessionId as required by the patched interface
- If upgrading is not immediately possible, disable the experimental tasks feature or isolate each session to its own InMemoryTaskStore instance
Patch Information
The fix is delivered in MCP TypeScript SDK 1.32.0. See the GitHub Security Advisory GHSA-22jm-h49p-29qw, Pull Request #2925, and commit 727075b for the full patch. The interface change requires custom TaskStore implementations to accept and enforce sessionId on every method.
Workarounds
- Instantiate a dedicated InMemoryTaskStore per session instead of sharing one across sessions
- Disable the experimental tasks feature on MCP servers until the SDK upgrade is deployed
- Restrict network access to the MCP server so only trusted clients can establish sessions
# Upgrade the MCP TypeScript SDK to the patched release
npm install @modelcontextprotocol/sdk@^1.32.0
# Verify the installed version
npm ls @modelcontextprotocol/sdk
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

