AI Governance Institute
← News
Research2026-11-18

Expired Domain Hijack Altered Kokoro AI Model Weights in 87 Seconds

What happened

Researchers Alon Hertz and his team published findings reported by Bleeping Computer. The findings show that Hugging Face does not verify whether a publisher account is still held by its original owner. The team registered the expired domain listed in Kokoro's public donation file, used it to reset the account password, and gained full control of the publisher account. They then ran a local test showing that Kokoro's neural weights could be modified in 87 seconds using a gaming graphics card. The model produced audio output that differed from the input text. All application code, file sizes, architecture, and delivery routes remained unchanged. The researchers confirmed that conventional malware scanners would find nothing to flag, because only the numeric weight values changed and no executable code was inserted. Hertz reported the issue to Hugging Face and received no response at the time of publication. The researchers did not upload modified weights to Hugging Face. Organizations including Google, Nvidia, AMD, Intel, and Amazon Web Services have published deployment guides or developer tool integrations that use Kokoro, illustrating how widely a poisoned update could propagate. The researchers noted the technique generalizes to any Hugging Face publisher account whose linked domain has lapsed.

Why it matters

  • ·Standard malware scanning and code-review processes cannot detect weight-level alterations. Organizations relying only on those controls have a blind spot that existing procurement or security frameworks do not yet close.
  • ·Enterprises using automated pipelines that fetch the latest version of an open-weight model without a manual review gate could silently receive a poisoned update. Audit trails tied to code changes would show nothing, breaking the traceability assumptions that underpin AI accountability programs and frameworks like the NIST AI Risk Management Framework (AI RMF 1.0) and Playbook.
  • ·The finding is not limited to Kokoro. Any Hugging Face publisher whose linked domain has expired is a potential takeover target, so organizations must assess every open-weight model in their inventory, not just this one title.

Stay ahead of stories like this

Get every Global AI governance development like this one, plus the rest of the week's developments. Every Thursday.

Free every Thursday. Unsubscribe anytime.

Governance controls affected

What to do now

  • ☐Inventory every open-weight or third-party AI model your organization pulls from Hugging Face or similar public repositories, and confirm whether any are configured to auto-update when the publisher pushes a new version.
  • ☐For each model in that inventory, check whether your intake process includes a step to verify the cryptographic hash or checksum of weight files against a known-good reference before the model is used in production.
  • ☐Ask your engineering or ML team whether the Kokoro text-to-speech model, or any model from the Hexgrad publisher account, is used in production, and confirm which version is currently deployed.
  • ☐Review your model update and change-approval process to confirm that updates to neural weight files trigger the same review gate as a code change, not just an automated download.
  • ☐Add publisher account integrity to your open-weight model due diligence checklist: verify that the domain and contact details associated with a Hugging Face publisher account are current and controlled by the legitimate owner before deployment.

What to watch next

Hugging Face has not publicly responded to the researchers' disclosure, so compliance teams should monitor whether the platform introduces publisher verification or domain-expiry checks. This finding adds weight to the emerging push for AI Bills of Materials. These would require model publishers to attest to the provenance and integrity of weight files, as covered in AI BOMs Are the Missing Link in Enterprise Supply Chain Transparency. Regulators developing supply chain requirements under frameworks such as the EU AI Act (Regulation (EU) 2024/1689) may treat model repository integrity as a conformity requirement for high-risk deployments. Teams should track any guidance from the EU AI Office on third-party model sourcing. Organizations should also watch whether the NIST AI Technology Evaluation Program issues criteria for weight-level integrity verification.

Related Coverage

Enforcement2026-11-17

New Commerce Rule and Blackwell Chip Findings Put Open-Source AI Supply Chains at Risk

A new US Department of Commerce rule announced November 16, 2026, targets remote access to Nvidia chips via third-country data centers. This workaround has powered China's open-source AI models. Separately, unverified research by former OpenAI researcher Will DePue claims Chinese AI companies had substantially greater access to restricted Nvidia Blackwell chips than previously estimated. Together, these developments create new export control compliance questions for enterprises that source open-weight models trained in China.

Research2026-11-16

Alpha School AI Misses Math Target for Half of K-8 Students

Internal data from Alpha School show that only 42% of students met the network's core promise of learning twice as fast in both math and reading. Former employees report the AI generated faulty lesson plans acknowledged internally to sometimes do 'more harm than good,' and that the company scraped third-party course content without permission. Alpha's platform now reaches more than 2,600 public school students in Texas and South Carolina.

Research2026-11-14

Blackwell GPU Fingerprints Found in Six Chinese Frontier Model Weights

Forensic analysis by Unemployment Labs found numerical patterns in publicly released model weights from DeepSeek, Moonshot AI, MiniMax, and Ant Group. These patterns match NVIDIA Blackwell graphics processing unit (GPU) hardware, which requires a U.S. export license to ship to China. Six models released since April 2026 carry Blackwell fingerprints, while earlier models from the same labs match the older Hopper generation. The researchers explicitly note they cannot determine where the chips are located, how access was obtained, or whether export regulations were violated.