top of page

Researchers Took Over the Hugging Face Account Behind Kokoro by Registering a Domain Nobody Owned. Your Code Pulls Whatever Is on main. Pin the Commit Hash Tonight

Writer: Patrick Duggan
Patrick Duggan
1 hour ago
6 min read

If anything you run loads the Kokoro text-to-speech model, pin it to a commit hash today. The repo is hexgrad/Kokoro-82M, and its current commit is f3ff3571791e39611d31c381e3a41a3af07b4987, last changed Thursday, April 10, 2025 (read from the Hugging Face API on November 18, 2026). Then do the same for every other model your agents pull. Over the past weekend, researchers showed that the account publishing a model with about 19 million pulls in the last month could be taken with a domain registration and a password reset. The official kokoro Python package fetches whatever is on the main branch every time it needs a file it has not already cached.


Nothing malicious was served. The researchers say they changed nothing and uploaded nothing, and the repo's last commit is still the April 2025 one. The account was the only weak point: no zero-day, no phishing, no breach of Hugging Face.



What happened


Security researcher Alon Hertz published the write-up on Medium on Monday, November 16, 2026, under the title "One click from 180 million downloads." His account of the domain: "The domain matching it — had no owner." And: "Anyone could register it, so we did." BleepingComputer, which covered it on November 17, adds that the domain was hexgrad.com and that it appeared in the repo's own DONATE.md file. With the domain registered, they requested a password reset for the publisher's Hugging Face account, received it, and controlled the account.


What that account reaches, in the researcher's numbers: "roughly 180 million all-time downloads," "143.5 million downloads from its main Hugging Face repository alone," and "about 19 million pulls last month alone." Those are his figures across the Kokoro ecosystem; we have not checked them. What we did check: on November 18 the Hugging Face API reported 13,092,988 downloads for hexgrad/Kokoro-82M, which is the Hub's rolling 30-day count for that one repo. BleepingComputer's "nearly 13 million monthly downloads" matches that figure. The write-up also lists projects that use Kokoro, including Google's AI Edge Gallery, NVIDIA NeMo voice agents, AMD GAIA and Intel OpenVINO; we have not checked those.



What an attacker could have pushed


The researchers did not demonstrate code execution. They demonstrated something quieter. They fine-tuned Kokoro's weights so that, with the inference code unchanged, text about a cake recipe came out as speech about an omelette. That took 87 seconds on a gaming GPU. In a second test they used a small LoRA fine-tune on Qwen2.5-0.5B-Instruct and got the planted answer in 80 of 81 targeted prompts, against 0 of 30 for the stock model. The modified weights were built and tested locally and never uploaded. BleepingComputer reports that a test from a separate account showed Hugging Face would not flag such an update as suspicious.


That is the important point for anyone building agents. A new revision with changed weights is the same file format, the same size and the same filename. A malware scanner has nothing to find. Your agent simply starts saying something different.


The louder option was also there. We read the kokoro package source on GitHub (main branch, November 18, 2026): the weights file is kokoro-v1_0.pth and the 54 voice files are .pt. Both are PyTorch pickle formats, not safetensors. The package downloads them with hf_hub_download and no revision argument, so it follows main. It does load them with torch.load(..., weights_only=True), which blocks ordinary pickle code execution, but only on a current PyTorch. CVE-2025-32434 (GHSA-53q9-r3pm-6pq6, published April 17, 2025) is remote code execution through torch.load with weights_only=True in every PyTorch before 2.6.0. So on an old torch, a hostile .pth pushed to main could have run code on the machine loading it. We did not test this, and nobody claims it happened.



Am I affected? Check what you actually downloaded


  1. Find it in the cache. By default the Hugging Face cache is ~/.cache/huggingface/hub (or wherever HF_HOME or HF_HUB_CACHE points). Kokoro is in models--hexgrad--Kokoro-82M. Inside, the file refs/main holds the commit hash your machine last resolved for main, and snapshots/ has one folder per commit you have ever pulled.

  2. Compare the hash. Run cat ~/.cache/huggingface/hub/models--hexgrad--Kokoro-82M/refs/main and ls ~/.cache/huggingface/hub/models--hexgrad--Kokoro-82M/snapshots/. If you see only f3ff3571791e39611d31c381e3a41a3af07b4987, you have the April 2025 files. A folder with any other hash is one you should explain.

  3. Check everything else. Run huggingface-cli scan-cache (newer huggingface_hub releases also offer it under the hf cache commands). It lists every cached repo with its revisions. Any repo without a commit hash in your code or lockfile is following someone else's main branch.

  4. Check containers and CI. Images that download at build time hold whatever main was on build day, and runners with an empty cache fetch it again on every run. Search your Dockerfiles and CI for from_pretrained, hf_hub_download, snapshot_download and KPipeline, and look for a revision argument.

  5. Check torch. Run python -c "import torch; print(torch.__version__)" in every environment that loads .pth or .pt files. Anything below 2.6.0 is exposed to CVE-2025-32434 whatever the model's provenance.


What to change


  • Pin by full commit hash, not by branch or tag. hf_hub_download, snapshot_download and from_pretrained all take revision="<40-character sha>". A tag can be moved by whoever holds the account, and that is exactly what was taken here.

  • Kokoro specifically: fetch the pinned files yourself, then pass local paths. KModel(config="path/config.json", model="path/kokoro-v1_0.pth") skips the download, and KPipeline loads a voice from a local path when the name ends in .pt.

  • Run production offline. Set HF_HUB_OFFLINE=1 so the runtime uses only what is already in the cache and never quietly resolves main again.

  • Mirror internally. Copy approved model revisions to storage you control, record their hashes, and have production read only from there. Changing a model then goes through code review instead of happening the next time someone else's account is used.

  • Never use trust_remote_code=True on an unpinned repo. That flag runs Python from the repo itself. Kokoro does not need it, but plenty of models do. With it set and no pinned revision, whoever controls the account controls your process.

  • Prefer safetensors and keep torch current. Safetensors files hold data only and cannot run code when they load. Where only pickle formats exist, use weights_only=True on PyTorch 2.6.0 or later.

  • Check the domains behind your own publishing accounts. The weak point was an email domain linked to a publisher that had gone quiet. If your team publishes models or packages, make sure the domain on every recovery email is registered, renewed and on auto-renew, and turn on two-factor authentication so a password reset alone is not enough.


What is confirmed, and what is not


Confirmed by the primary write-up: the domain was unregistered and the researchers registered it; the account was taken by password reset; the weight-tampering demos; "Nothing was leaked, changed, or uploaded, and the account has been secured"; "We are coordinating with Hugging Face." Confirmed by our own reads: the repo's commit and its April 10, 2025 timestamp, the 30-day download count, the file formats, and how the kokoro package loads them (unpinned downloads, weights_only=True).


Not confirmed: the write-up does not say who secured the account or when. BleepingComputer says neither Hertz nor BleepingComputer had heard back from Hugging Face. We found no statement from Hugging Face or from hexgrad in the sources we read, and hexgrad's Hugging Face activity page shows no recent posts. The write-up does not mention two-factor authentication, PyPI or the GitHub repo. Whether those accounts used the same domain for recovery is not established. The write-up gives no disclosure dates. We read it through the author's Medium feed, because the article page returned 403 to our fetcher.


Indicators: none published in the sources we read, and none apply. This was research, and the model repo is legitimate, so we are adding nothing to the feed.



How we know


We read the researcher's write-up, the BleepingComputer article, the Hugging Face model API for hexgrad/Kokoro-82M, the kokoro package source (kokoro/model.py and kokoro/pipeline.py on GitHub, main branch) and the GitHub advisory for CVE-2025-32434. We did not download, open or load any file from the model repo.



Sources


Read next: Microsoft Researchers Backdoored an AI Agent's Model by Editing Seven Neurons. Treat Model Files Like Code, the same lesson from the other side: a few changed weights, and the files still look normal.




Every indicator in this post is in the feed. Free.

1.58M+ IOCs, STIX 2.1 / TAXII, 88% novel vs ThreatFox, exploited-CVE leads ahead of CISA. No credit card — a free API key in 30 seconds, and you can audit every claim above against the live endpoints.


Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page