"We changed what this model says by fine-tuning its neural weights. Same input text. Same voice file. Same configuration. Same inference code — not a single line touched in the application that runs it," said security analyst Alon Hertz.
Alon Hertz’s demonstration: silent manipulation by changing only model weights
Hertz and his research team showed that an attacker who controls a publisher account can alter a text-to-speech model’s outputs by adjusting only the model’s weights. In a controlled test they modified Kokoro’s weights in 87 seconds using a gaming GPU so the model produced "poisoned" audio — while keeping the inference pipeline, voice files, and application code unchanged. Because the change was limited to neural weights, the researchers say there was no malicious code or signature for malware scanners to detect.
How the Hugging Face account was taken over
The researchers gained control of Kokoro’s Hugging Face publisher account by registering the "hexgrad.com" domain listed in the model’s public DONATE.md file and using it to perform a password reset. Over the past weekend the team completed the takeover, then tested their modified weights locally; they did not upload those weights to Hugging Face, citing the risk to production systems that depend on the model. Hertz reported the issue to Hugging Face but, he says, received no response. BleepingComputer also contacted Hugging Face for comment and has not received a response yet.

Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleKokoro’s footprint: downloads, silence from the developer, and third‑party tooling
Kokoro is an open-weight, multilingual text-to-speech model with more than 50 voice presets. Although its developer, Hexgrad, has not pushed an update since April 2025, the model remains widely used: the project has nearly 13 million monthly downloads on Hugging Face. The Kokoro GitHub repository and the Python package on PyPI also went silent around the same time Hexgrad stopped releasing updates. The model is supported by developer tools from Google, NVIDIA, and Intel, and AMD and Amazon Web Services have published guides demonstrating Kokoro’s use in AI-agent workflows.
Supply-chain risk: legitimate-looking updates as a delivery channel
The researchers warn that compromised publisher accounts could be used as distribution channels in supply-chain attacks. A modified release would look like a legitimate update — similar publisher profile, similar file routes, file format and architecture, and comparable file sizes — and therefore would be unlikely to trigger suspicion or malware scanners. Hertz notes that Hugging Face does not vet publishers or verify whether an account is still held by its original owner, and that unvetted weights and datasets "flow through mirrors and converted versions, GitHub releases, package registries, deployment guides, and direct downloads across the entire open-model ecosystem." Tests from another account, the team says, showed that Hugging Face would not flag such a release as suspicious.
How Google, NVIDIA, Intel, AMD, and Amazon Web Services are implicated
- Google, NVIDIA, and Intel: These developer-tool partners are named in the record as providers of software that supports Kokoro. Their tooling can be part of developers’ normal build and deployment pipelines that pull in open models like Kokoro, which means altered weights distributed through a compromised publisher account could reach systems that use those toolchains.
- AMD and Amazon Web Services: Both vendors have published guides demonstrating Kokoro's use in AI-agent workflows, linking their documentation and users to the model; any widely distributed weight change could therefore propagate through published examples and orchestrated agent setups.
- Enterprises and deployment teams using Kokoro: Organizations that automatically fetch new revisions or follow published guides could receive a malicious update that appears legitimate, because the release would contain no executable malware and would not alter the surrounding application code.
Hertz framed Kokoro as a high-profile example used to highlight a broader trust problem: accounts and content on open-model platforms can be curated for a period and then changed, with downstream systems trusting routine updates. He says he alerted Hugging Face; at the time of reporting, the platform had not replied. The test stopped short of publishing modified weights precisely because doing so would likely endanger production systems — a restraint that nonetheless underscores the practical risk: a stealthy, weight-only modification can alter what a deployed AI says without leaving the telltale traces defenders currently rely on.
Read the original BleepingComputer story: https://www.bleepingcomputer.com/news/security/researchers-hijack-hugging-face-account-for-popular-kokoro-ai-model/




