/

CVE-2026-73498

Severity:

HIGH RISK

(7.7)

Path Traversal (CWE-22) in MCP Atlassian Attachments

CVE-2026-73498

Deferred

Vulnerability type

Category:

Path Traversal

CWE:

CWE-22

Impact:

Sensitive File Disclosure and Credential Exposure

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 an authenticated MCP client to provide a file path that is opened by the attachment upload workflow without sufficient path validation. An attacker may read files accessible to the server process and exfiltrate their contents to Confluence, including credentials or environment variables when those files are readable.

Affected context

The issue occurs in MCP Atlassian deployments where the Confluence attachment upload tool accepts a client-controlled file path and passes it to the file-opening operation without applying the expected safe-path validation. Risk is higher when the MCP server process can read sensitive host or container files, or when an AI agent can be influenced by untrusted content to invoke the tool.

Exploit preconditions

  • Attack surface: The MCP Confluence attachment upload tool
  • Required access: An authenticated MCP client, or an indirect ability to cause an AI agent to invoke the tool
  • Trigger point: Processing a client-supplied path during attachment upload
  • Why exploitation is possible: The supplied path reaches file access logic without sufficient confinement to an approved directory
  • Additional condition: The targeted file must be readable by the MCP server process

Why this happens

The vulnerability results from treating a client-controlled file path as trusted input and passing it directly to a file-opening API. Without canonicalization, allowlisting, and enforcement that the resolved path remains within an intended storage directory, the operation can access files outside the permitted attachment area. This behavior maps to CWE-22: Improper Limitation of a Pathname to a Restricted Directory, commonly known as path traversal.

Potential impact

  • Sensitive local files may be disclosed
  • API tokens, credentials, and environment variables could be exfiltrated
  • Confidential application or host configuration may be exposed
  • Stolen credentials may enable access to connected Atlassian or cloud services
  • An AI-driven workflow may be manipulated into performing unintended file disclosure

General mitigation guidance

Mitigation generally focuses on reducing trust in client-provided paths and enforcing a narrow file-access boundary.

Common defensive patterns include:

  • Resolving and canonicalizing paths before access
  • Enforcing that resolved paths remain within a dedicated attachment directory
  • Preferencing opaque server-side file identifiers over arbitrary client paths
  • Applying strict allowlists for permitted files and extensions where appropriate
  • Rejecting symbolic-link escapes and other filesystem boundary bypasses
  • Running the MCP server with least-privilege filesystem permissions
  • Keeping secrets out of broadly readable environment files and mounted paths
  • Requiring explicit approval for agent-initiated file-upload operations

How teams detect issues like this

Organizations can identify path traversal risks through:

  • Static analysis rules that trace user-controlled path values into file APIs
  • Code review checks for direct use of client input with open, read, archive, or upload functions
  • Tests that verify canonical paths cannot escape approved directories
  • Runtime monitoring of attachment requests involving unexpected directories or sensitive file locations
  • Policy enforcement requiring centralized safe-path validation for file access tools
  • Automated code review tools that flag missing path confinement and authorization checks

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-g5r6-gv6m-f5jv

Frequently asked

questions

It can be exploited through the MCP interface when an authenticated client can invoke the attachment upload tool and the server can read the selected file. An AI agent may also become an indirect trigger when untrusted content influences tool use.

The attacker does not necessarily need direct shell or filesystem access. They need a permitted MCP interaction that causes the server to open a file readable by its process.

Restrict file access to a dedicated approved directory, canonicalize paths, reject escapes and symlinks as appropriate, and use server-side file identifiers instead of arbitrary client-supplied paths.

Apply centralized path-validation policies, use least-privilege service accounts, add boundary-focused tests, monitor unusual file access, and require explicit authorization for agent-driven file operations.

Find out what's already
exploitable in your codebase.