Home / Blog / Researchers hijack Hugging Face account for popular Kokoro…
Tech News

Researchers hijack Hugging Face account for popular Kokoro AI model

Researchers hijacked the Hugging Face account for the Kokoro speech model via an expired domain, exposing millions of systems to poisoned weights.

By Dillip Chowdary • Nov 18, 2026 • Source: BleepingComputer

Researchers hijack Hugging Face account for popular Kokoro AI model
Image: BleepingComputer

At a glance

Product
Kokoro open-weight text-to-speech model
Developer
Hexgrad
Monthly downloads
Nearly 13 million downloads
Poisoning time
87 seconds on a gaming GPU
Last developer update
April 2025
Platform affected
Hugging Face

Security researchers took control of the Hugging Face account hosting the popular open-weight text-to-speech model Kokoro by exploiting absent publisher identity checks, as detailed in BleepingComputer's report. The platform takeover allowed researchers to seize the official distribution channel of an artificial intelligence model that logs nearly 13 million monthly downloads. While automated supply-chain takeovers often deliver code-execution exploits, this demonstration replaced neural weights directly, altering spoken output without triggering signature scanners or modifying application logic.

This report examines how the takeover occurred, how weight poisoning evades enterprise malware defenses, and the exposure across major cloud and hardware ecosystems. It is intended for machine learning engineers, security teams, and infrastructure architects who automatically pull open-weight models into production speech pipelines.

01Domain recovery and account password reset

The takeover occurred because the platform lacked basic verification to ensure an account remains in the hands of its original owner. Developer Hexgrad had left the project unmaintained since April 2025, leaving the primary GitHub repository and the Python package on PyPI dormant. The development team had also allowed the registration of its official domain name to lapse, leaving an exploitable gap in account recovery systems.

Security researcher Alon Hertz and his team inspected the public DONATE.md file inside the Kokoro repository to find the publisher's contact address. After discovering that hexgrad.com had expired, the researchers purchased the domain over the weekend. They subsequently requested a password reset on the Hugging Face publisher account, received the recovery email on their newly acquired domain, and gained full administrative control of the repository.

  1. Discover dormant domain Inspect the public DONATE.md file to locate the publisher contact domain.
  2. Register lapsed asset Purchase the abandoned hexgrad.com domain name through a public domain registrar.
  3. Trigger password reset Submit an automated password reset request on the official Hugging Face account.
  4. Assume repository control Intercept the platform recovery email to establish administrative rights over model releases.

02Model weight poisoning on consumer hardware

Rather than using account access to plant executable malware or backdoors in application code, the researchers evaluated weight modification. Using a standard consumer gaming GPU, the team fine-tuned Kokoro's neural weights in 87 seconds. The targeted adjustment caused the speech-synthesis model to utter audio phrases that deviated entirely from the provided text inputs.

Because standard security scanners look for executable files, unusual system calls, or known binary signatures, altered neural weights pass through without inspection. The modified weights preserved the original file format, file size, and neural architecture. Software running the model operated normally because the underlying inference code, audio presets, and configuration parameters remained untouched.

Note: The researchers executed all fine-tuning and inference verification inside an isolated local testing environment, deliberately withholding poisoned weights from Hugging Face to avoid disrupting live enterprise integrations.

Advertisement

Tech Pulse Daily

Get tomorrow's pulse first

Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.

03Integration across developer tools and platforms

Kokoro functions as an open-weight engine converting text into speech across multiple languages with more than 50 voice presets. Its footprint spans foundational frameworks distributed by major hardware and cloud corporations. Google, NVIDIA, and Intel support the model within dedicated developer tools designed for autonomous voice agents and on-device speech synthesis.

AMD and Amazon Web Services maintain deployment guides that instruct engineers on incorporating the model directly into enterprise agentic workflows. Because these architectures often pull weights automatically through scripts, a poisoned update distributed from the verified publisher handle would flow directly into live client applications without raising security flags.

AttributeKokoro Text-to-Speech Model
DeveloperHexgrad
Monthly downloadsNearly 13 million
Last updateApril 2025
Voice presetsMore than 50 presets
Hardware partnersGoogle, NVIDIA, AMD, Intel, AWS

04Supply-chain vulnerabilities in open model ecosystems

The incident highlights structural gaps across open-source artificial intelligence distribution channels, where identity verification remains absent. Model hubs operate similarly to early code registries, treating an account name as an implicit trust token without validating ownership continuity. If a maintainer abandons a service domain or relinquishes email hosting, third parties can easily re-register those communication channels.

A compromised account can silently distribute poisoned model revisions to downstream pipelines that pull floating release tags. The threat extends beyond compromised maintainers to malicious actors who deliberately maintain legitimate-looking repositories over time before shipping compromised checkpoints. Automated ingestion pipelines that blindly fetch updated neural files remain completely vulnerable to silent data manipulation.

05Reporting delays and vendor notification silence

Following the successful takeover, Hertz submitted a vulnerability disclosure to Hugging Face detailing the verification gap and the potential for weight manipulation. The platform did not issue an immediate response to the researcher's disclosure. Independent inquiries sent by journalists regarding publisher verification policies also received no reply at the time of publication.

Without platform-level domain re-verification, downstream systems must implement internal validation controls before pulling model updates into staging or production. Model consumers face risks when treating model hubs as verified package registries, especially when models continue to rack up millions of monthly downloads after their original creators have stopped maintaining the underlying infrastructure.

Frequently asked questions

How did the researchers hijack the Kokoro account on Hugging Face?

The researchers registered the expired hexgrad.com domain listed in the model repository DONATE.md file and initiated a standard account password reset.

What happens when an attacker modifies a model's weights instead of application code?

The model produces audio that diverges from the input text while bypassing malware scanners, because the file structure, code, and system calls remain clean.

Which major companies provide tools or workflows supporting Kokoro?

Developer tools and implementation guides supporting Kokoro are provided by Google, NVIDIA, AMD, Intel, and Amazon Web Services.

Sources

Dillip Chowdary

Author

Dillip Chowdary

Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.

Related on Tech Bytes

Advertisement

5-min tech signal

Weekday briefing for engineers who skip the noise.

No spam · Unsubscribe anytime

Advertisement

✈️ CareerPilot

Your AI job-search copilot

Match your resume against live Ashby, Greenhouse & Lever openings — fit scores, job-specific resume optimization and email alerts.

Find matching jobs →

Free Tools

Browse all tools →