Summary
- Researchers obtained the Kokoro model publisher's account by registering an expired domain associated with its recovery identity.
- A controlled experiment demonstrated that altered model weights could change generated speech without modifying the application.
- No malicious model update was publicly distributed, but the findings expose weaknesses in publisher identity and AI artefact assurance.
Researchers gained control of the publishing account behind a widely used artificial intelligence speech model, exposing how an abandoned domain could undermine trust in software distributed through a recognised AI repository.
The investigation involved Kokoro, an open-weight text-to-speech model hosted on Hugging Face. Security researcher Alon Hertz and colleagues registered a domain previously associated with the model’s publisher, then used control of that domain to recover access to the publishing account.
The researchers subsequently demonstrated that modifying the model’s numerical weights could change its output without altering the application code responsible for loading and running it. They did not distribute their modified model publicly, and the disclosure does not establish any downstream criminal compromise.
The findings were reported on 17 November 2026 and concern two related weaknesses: trust in the identity distributing an AI model, and the possibility of manipulating model behaviour while leaving conventional application code unchanged.
Abandoned domain enabled account recovery
Hugging Face operates a widely used platform through which researchers, developers, and businesses obtain AI models and associated development resources.
Model publishers can accumulate considerable trust as their repositories attract downloads and become integrated into commercial systems. Developers may subsequently retrieve updated model files using the same publisher identity without independently examining every change.
In the Kokoro investigation, the researchers identified the domain hexgrad.com in publicly available project information. Although the domain remained associated with the model’s publisher, it was available for registration.
The team acquired the domain and used the associated email identity to complete a password recovery process for the Hugging Face account. This gave the researchers control of the publisher account rather than simply access to an unrelated website.
The technique did not require breaking the platform’s authentication software or stealing the publisher’s existing password. It depended on the relationship between an account recovery address and an external domain whose ownership had lapsed.
That relationship can persist for years after a project is created. Organisations may change personnel, abandon websites, or discontinue email domains without removing references to those identities from development accounts.
Where an abandoned identity retains recovery privileges, a new domain owner may obtain access that the original publisher never intended to transfer.
Altered weights changed model behaviour
After demonstrating account takeover, the researchers examined the consequences of a malicious publisher introducing modified model files.
In a controlled experiment using a consumer graphics processor, they changed Kokoro’s weights in approximately 87 seconds, causing the model to produce spoken words that differed from the text it had been given.
The surrounding inference code remained unchanged. The manipulation involved the numerical parameters that determine how the model processes inputs and generates output.
This distinguishes model manipulation from a conventional application compromise involving an altered executable or an explicitly malicious software dependency.
AI models can contain millions or billions of numerical parameters. Those values influence behaviour, but inspecting them does not provide the same straightforward account of functionality as reviewing conventional program instructions.
As a result, a model may retain its expected file format and architecture while producing outputs that differ from the legitimate version.
The experiment demonstrated technical feasibility rather than an observed criminal attack. The researchers did not upload manipulated weights to the public Kokoro repository, and no affected production organisation was identified in connection with their test.
Model provenance becomes an operational requirement
Organisations increasingly incorporate externally maintained AI models into applications without controlling the accounts through which those models are distributed.
That arrangement introduces dependencies involving publisher identity, repository access, version management, file integrity, and the processes used to approve updates.
Pinning a known model revision can help prevent an application from automatically accepting subsequent changes. However, pinning does not establish that the original artefact was trustworthy, nor does it address the possibility that a publisher’s account was compromised before the selected version was released.
Verification procedures therefore need to consider the model’s provenance alongside the integrity of the files being deployed. A cryptographic hash can establish whether an artefact has changed, but its value as an assurance control depends on whether the original reference was obtained through a trusted process.
Account recovery arrangements deserve similar attention. Domains, email addresses, and administrative identities associated with important software repositories can remain security dependencies long after their original operational purpose has changed.
For AI models supporting customer-facing or operational processes, unauthorised changes could potentially affect the reliability of information or actions generated by dependent applications. The consequences would depend on the model’s function, its permissions, and how organisations use its output.
The Kokoro investigation does not establish that such consequences occurred. It demonstrates how account ownership and model integrity can become connected through a trusted distribution channel.
The researchers reported coordinating the findings with the relevant parties. Their work provides a controlled example of how a weakness in conventional account management could create opportunities to manipulate AI components relied upon by downstream developers.




