/

CVE-2026-73496

Severity:

HIGH RISK

(7.7)

Path Traversal (CWE-22) in MCP Atlassian Uploads

CVE-2026-73496

Deferred

Vulnerability type

Category:

Path Traversal

CWE:

CWE-22, CWE-73

Impact:

Sensitive Data Disclosure

Severity:

HIGH RISK

(7.7)

CVSS version:

3.1

Catch this CVE while code is being written

Detect vulnerable code directly in pull requests, so CVEs are fixed before merge.

Summary

MCP Atlassian allows client-controlled attachment paths to be resolved on the MCP server without restricting them to an approved workspace. In remote HTTP, SSE, or multi-user deployments, a client with write-tool access may cause server files, credentials, or other tenant data to be uploaded to Atlassian.

Affected context

The vulnerability occurs in deployments where MCP Atlassian handles file-upload requests on behalf of remote or multiple users. Risk is present when attachment tools accept absolute or traversal-based paths and the server process can read files beyond an intended workspace; a local single-user stdio deployment generally does not introduce the same trust boundary.

Exploit preconditions

  • Attack surface: Confluence and Jira attachment-upload tools exposed through MCP
  • Required access: Client access to the relevant write tools
  • Trigger point: Processing an attachment path during an upload request
  • Why exploitation is possible: User-controlled paths are resolved without confinement to an approved server workspace, allowing traversal or absolute paths to reference readable files

Why this happens

The underlying bug is insufficient filesystem path validation in file-upload handling. Client-provided paths are treated as usable filesystem locations without canonicalizing them and verifying that the resolved path remains within an explicitly trusted directory. This maps primarily to CWE-22: Improper Limitation of a Pathname to a Restricted Directory, with CWE-73: External Control of File Name or Path also describing the failure to constrain externally supplied path values.

Potential impact

  • Sensitive server files may be disclosed to an Atlassian instance
  • Environment-held Atlassian credentials or configuration secrets could be exposed
  • Data belonging to another tenant may be uploaded or disclosed
  • The MCP server's filesystem confidentiality could be reduced
  • Organizations could face cross-tenant data exposure in shared deployments

General mitigation guidance

Mitigation generally focuses on reducing trust in client-supplied paths and enforcing a narrow filesystem boundary.

Common defensive patterns include:

  • Restricting uploads to a dedicated, approved workspace
  • Canonicalizing paths before access and verifying containment within that workspace
  • Rejecting absolute paths, traversal components, symbolic-link escapes, and unexpected path forms
  • Prefering server-generated file identifiers or uploaded content streams over arbitrary filesystem paths
  • Running the MCP service with least-privilege filesystem and network permissions
  • Separating tenant workspaces and credentials in multi-user deployments
  • Avoiding storage of secrets or broadly readable sensitive files in upload-accessible locations

How teams detect issues like this

Organizations can look for this issue through:

  • Static analysis of file APIs receiving values from MCP tool arguments
  • Code review checks for missing canonicalization and workspace-containment validation
  • Tests covering absolute paths, parent-directory traversal, symbolic links, and cross-tenant workspace access
  • Policy enforcement requiring all upload paths to originate from approved directories or opaque server-side handles
  • Runtime monitoring for attachment requests that reference unexpected directories or sensitive file locations
  • Automated security review tools that flag externally controlled path arguments reaching filesystem operations

Sources

SourceLink
github.comgithub.com/sooperset/mcp-atlassian/commit/b041733473f95119dd539542a43c280737a8e460
github.comgithub.com/sooperset/mcp-atlassian/pull/1448
github.comgithub.com/sooperset/mcp-atlassian/releases/tag/v0.22.0
github.comgithub.com/sooperset/mcp-atlassian/security/advisories/GHSA-wm45-qh3g-v83f

Frequently asked

questions

It can be exploitable in remote HTTP, SSE, or multi-user deployments when a client has access to the affected write tools. A local single-user stdio deployment generally does not cross the same server-side trust boundary.

The attacker needs client access that permits use of the Confluence or Jira attachment-upload functionality. The server process must also be able to read the targeted file.

Constrain all uploads to a dedicated approved workspace after canonicalization and containment checks. Server-generated file references or controlled content streams are safer than accepting arbitrary client filesystem paths.

Apply least-privilege filesystem permissions, isolate tenant storage, reject traversal and symbolic-link escapes, and test every file path that originates from an external tool argument.

Find out what's already
exploitable in your codebase.