Loading vulnerability…
CVE-2026-5843 is an arbitrary code execution vulnerability in the MLX inference backend used by Docker Model Runner on macOS. The issue arises because the MLX-LM library processes the model_file field in a model's config.json by unconditionally importing and executing the referenced Python file through Python's import mechanisms, without a trust gate or equivalent validation. If a malicious model package specifies a Python file in that field, the file is executed during model loading. Because the MLX backend runs without sandboxing, the imported code executes on the Docker host in the security context of the Docker Desktop user rather than being confined to an isolated environment.
Mallory correlates every CVE against your assets, your vendors, and active adversary campaigns. Know which vulnerabilities matter for you, not just which ones are loud.
What it means. What to do now. Patch path, mitigations, and the assume-compromise checklist.
What an attacker gets, and what they’ve been doing with it.
If you can’t patch tonight, do this now.
Patch, then assume compromise.
1 valid exploit after Mallory filtered fakes, detection scripts, and README-only repos.
This repository is a compact proof-of-concept exploit for CVE-2026-5843. It contains two files: a README with usage instructions and one Python script, poc_cve_2026_5843.py, which implements the full exploit. The exploit stands up a minimal attacker-controlled OCI registry over HTTP on port 5555 and serves a malicious model artifact tailored for Docker Model Runner's mlx-lm loading path. The core capability is container-to-host remote code execution. The script builds a fake model package consisting of config.json, model.safetensors, and a malicious model.py. The config explicitly sets "model_file": "model.py", causing the target to import attacker-supplied Python during model loading. When a container instructs model-runner.docker.internal to pull and then run inference on the model, the host-side Model Runner imports model.py and executes the embedded Python payload as the logged-in host user. The payload is operational but simple: it creates ~/Desktop/mlx.txt and writes hostname, current user, id output, and timestamp. After that, it defines minimal ModelArgs and Model classes using mlx.nn so the artifact resembles a valid model and can proceed through the expected loading path. The script also generates a minimal safetensors blob with zero-filled arrays and constructs OCI manifest/config metadata, including layer annotations such as org.cncf.model.filepath, to make the registry responses look like a valid model distribution. Repository structure is straightforward: README.md documents prerequisites, affected versions, and curl commands to trigger the vulnerable pull and inference endpoints; poc_cve_2026_5843.py contains all exploit logic, including digest generation, fake model construction, blob storage, and an HTTP request handler implementing enough of the OCI Registry v2 API to satisfy the target. This is a real exploit PoC rather than a scanner or detector.
Products and vendors Mallory has correlated with this vulnerability. Open in Mallory to drill down to specific CPE configurations and version ranges.
Vendor-confirmed product mapping. Mallory continuously reconciles this list against your asset inventory.
5 sources tracked across advisories and community write-ups. News coverage will land here when it surfaces.
No news coverage yet. Advisories and community discussion only.
Query your assets running an affected version, and investigate the blast radius.
Every observed campaign linking this CVE to a named adversary.
Malware families riding this exploit, with evidence and IOCs.
YARA, Sigma, Snort, and vendor rules, auto-deployed to your SIEM.
Cross-references every affected SKU, including bundled OEM variants.
Community discussion across Reddit, Mastodon, and other social sources.