{
  "id": "kokoro-publisher-account",
  "url": "https://ohshit.app/incidents/kokoro-publisher-account/",
  "title": "Researchers take over Kokoro's Hugging Face publisher account through domain registration",
  "level": 3,
  "level_name": "sharted",
  "date": "2026-11-16",
  "date_is_approximate": false,
  "updated": null,
  "tags": [
    "AI",
    "supply-chain",
    "account-takeover"
  ],
  "ecosystems": [
    "hugging-face"
  ],
  "summary": "Alon Hertz reports taking over the hexgrad publisher account by registering hexgrad.com and resetting its password. The team tested altered Kokoro weights locally and says it secured the account without publishing them.",
  "splash": "Deployments that fetch Kokoro weights from hexgrad/Kokoro-82M and model publishers using domain-based recovery email.",
  "wipe": "Pin reviewed model revisions, audit recent downloads and protect publisher recovery email and 2FA.",
  "in_the_wild": "no",
  "source_exception": null,
  "cves": [],
  "fixed": [],
  "checklist": [
    {
      "title": "Pin the model revision used in production",
      "hint": "Inventory direct and transitive Kokoro model downloads. With hf_hub_download or snapshot_download, set revision to a reviewed full-length commit hash, rather than relying on the default main branch. Retain the approved weights and their hashes for comparison."
    },
    {
      "title": "Check what your deployments actually fetched",
      "hint": "Compare recent build records, model revision IDs and cached snapshots with your approved artifacts. Hold unreviewed model updates for evaluation of expected text-to-speech behavior before promoting them. Escalate unexpected file changes through your normal security response."
    },
    {
      "title": "Protect publisher recovery and sign-in",
      "hint": "Keep recovery email on a domain you control and renew it. Review account email and access settings. Enable Hugging Face 2FA under Settings > Security, store recovery codes securely and verify publisher access after account recovery."
    }
  ],
  "sources": [
    {
      "name": "Alon Hertz's original disclosure",
      "url": "https://medium.com/@alonhertz1/one-click-from-180-million-downloads-taking-over-the-hugging-face-account-behind-an-ai-model-ac61c16305b7",
      "says": "Describes the publisher-account takeover, local weight modification and account protection.",
      "primary": true
    },
    {
      "name": "Hugging Face download documentation",
      "url": "https://huggingface.co/docs/huggingface_hub/guides/download",
      "says": "Documents revision selection and full-length commit hashes for downloads.",
      "primary": true
    },
    {
      "name": "Hugging Face two-factor authentication guide",
      "url": "https://huggingface.co/docs/hub/security-2fa",
      "says": "Documents account 2FA setup and recovery codes.",
      "primary": true
    },
    {
      "name": "Researcher's public RSS feed",
      "url": "https://medium.com/feed/@alonhertz1",
      "says": "Dates the original article to November 16, 2026.",
      "primary": true
    }
  ],
  "headline": null,
  "one_liner": null,
  "in_the_wild_note": "Research demonstration. Hertz says the modified weights were never uploaded.",
  "mitigation": null,
  "timeline": [
    {
      "date": "Nov 16, 2026",
      "title": "Hertz publishes the account-takeover demonstration",
      "body": "The researchers say they secured the account and uploaded no modified weights."
    }
  ],
  "body_markdown": "Hertz says registering hexgrad.com enabled a password reset for the Hugging Face publisher account behind Kokoro. The researchers also tested altered model weights locally, causing speech to differ from its input. The disclosure says they coordinated with Hugging Face, secured the account and published no altered weights.\n\nThe demonstrated risk is a trusted publisher identity changing hands. This incident is rated L3 because it demonstrates a supply-chain weakness affecting a widely used model. Public evidence does not establish malicious distribution or downstream victim compromise. Pinning and evaluating model artifacts helps consumers control which revisions reach production."
}