MLflow: path traversal via model_id in the log outputs API (CVE-2026-127202)

A caller can use a crafted model identifier to write metadata outside the intended run directory on affected MLflow FileStore deployments. Upgrade to protect files writable by the service account.

Public CVE record · Vendor or publisher advisory · Native module source

CVE advisory record · Publisher advisory · Module source

Result. The marked request reached the backend in the baseline and raised its counter. The native rule refused that request with HTTP 403 and preserved HTTP 200 for the benign probe. The synthetic fixture ran through Eden 1.103.0 on November 12, 2026 (UTC).

The vulnerability

MLflow FileStore uses a model identifier when resolving where to persist run output metadata. In affected versions, an identifier with parent directory segments can make the server create directories and write metadata outside the expected run directory. The filesystem impact is bounded by the MLflow process’s permissions.

MLflow fixed the issue in 3.18.0. The advisory describes FileStore deployments. The CVE record reports a CVSS base score of 6.3.

The Eden patch

Tier 1 — declarative policy. The recorded policy denies requests to /api/v2/ai-models and its descendants. The empty method list leaves the deny unrestricted by HTTP method. This is a coarse rule and blocked the benign probe on the same path.

{
  "version": 1,
  "profile": "http-structure-v1",
  "policy": {
    "version": 1,
    "rules": [
      {
        "scope": {
          "methods": [],
          "path": "/api/v2/ai-models",
          "descendants": true
        },
        "require": {
          "kind": "deny"
        }
      }
    ]
  }
}

Tier 2 — native Rust module. gabriel_mlflowtrav_v1 waits for the request to be complete, checks whether the request target starts with /api/v2/ai-models, and looks for any configured marker in the body. The substring comparison is ASCII case-insensitive. Outside that prefix or without a marker, the module returns Continue. The shown lines are excerpts from the linked source.

const ROUTE: &[u8] = b"/api/v2/ai-models";
static NEEDLES: [&[u8]; 2] = [b"model_id", b"../"];

NEEDLES.iter().any(|n| find_ignore_case(body, n).is_some()) && return Ok(Decision::Refuse);

Implement the remediation

Use this procedure in an isolated fixture to reproduce the scoped restriction. Follow the vendor update for the underlying exposure. See deployment prerequisites.

1. Prepare the target. Verify the loaded product and version against the advisory. Use a compatible Eden HTTP interlay and an authenticated operator with Configure permission. Use the recorded declaration and the linked source when applicable.
2. Apply the declarative policy. Scope the policy to /api/v2/ai-models and descendants (all methods, deny). Set INTERLAY to the target interlay UUID in the authenticated Eden CLI. Run an authenticated GET /api/v1/interlays/{interlay} with X-Eden-Verbose: true. Save its complete current settings object as current-settings.json and retain the prior remediation declaration for recovery. Set IDEMPOTENCY_KEY to a fresh key for this intended mutation.

Save the JSON in the patch section as remediation.json.

jq -e --slurpfile policy remediation.json \
  '.remediation = $policy[0] | {settings: .}' current-settings.json \
  > interlay-update.json
eden api request PATCH "/api/v1/interlays/$INTERLAY" \
  --header "Idempotency-Key:$IDEMPOTENCY_KEY" \
  --body-file interlay-update.json --content-type application/json --confirm

3. Verify the serving policy. Read back the interlay.

eden api request GET "/api/v1/interlays/$INTERLAY" \
  --header 'X-Eden-Verbose:true'

Check that remediation_status.state is ready and both desired and effective SHA-256 values match the intended declaration. Replay the benign and marked requests in the recorded comparison on an isolated target. Check the recorded response and backend outcome for each probe. 4. Restore the previous policy. Coordinate the change with any other operator updating this interlay. Read current settings again and merge the saved remediation field into them. If the prior settings had no remediation field, omit it in the replacement. Send the same PATCH with a fresh idempotency key, then read back the intended state. Repeat the verification probes. Keep the affected route isolated if restoring the prior settings reopens exposure.

Vendor update. Use the advisory to select the fixed release for the installed branch. Reported fixed releases include 3.18.0. Follow the vendor’s update instructions and confirm the version actually loaded by the service.

Fixture and probes

The fixture was a synthetic HTTP handler behind Eden. It answered the probes and kept a separate counter that increased when the handler recognized its marker bytes. The counter measures a fixture effect. An HTTP 200 from the baseline does not by itself establish that the counter moved.

Request. POST /api/v2/ai-models/run for the benign probe and the marked probe.

The benign synthetic body supplied a training run and output label. The second body contained a model_id field with parent directory fragments. The backend counter recorded the matching bytes. No file was created by the fixture.

Measured comparison (2026-11-12 UTC)

The values after each status are the observed increases in the backend marker counter.

Request latency

Eden telemetry request spans measured in the fixture. Latencies are in microseconds (µs).

These are durations of the recorded request spans. The passing request includes the path to the fixture, while the policy and native refusal rows describe requests that ended at Eden. The rows represent different request outcomes, so their difference does not isolate module CPU time. The sample counts are shown for each arm.

Signed module authority

The module import returned HTTP 201 and required explicit activation. Signature verification was deferred at import. Read-only inspection returned HTTP 403, admission with a tampered statement returned HTTP 400, and admission with an invalid signature returned HTTP 403. The activation receipt reported the module serving on all expected shards.

After revocation, the same marked request received HTTP 503. Re-approval and reactivation restored HTTP 403. The source digest below identifies the linked Rust file. The binary digest identifies the artifact recorded in the activation evidence.

• Source SHA-256 726866334c147cbd6a822e673211012110b94ac19e6dd5ab96fffa3028e40ed4
• Tested binary SHA-256 a78f070f9908be9c009d6ee46a97a8c6312b387b2ff685b93c3e4eea05bb970a

Test scope

The fixture route is /api/v2/ai-models/run. The advisory’s POST /api/3.0/mlflow/runs/outputs route and FileStore writes are outside the tested scope. Either the model_id name or a traversal fragment can trigger the filter, so a normal model identifier can be refused on the fixture prefix. It does not resolve or normalize a filesystem path.

Evidence and artifacts

• Public CVE record
• Vendor or publisher advisory
• Published module source