Executive Summary
Researchers demonstrated a supply-chain compromise involving the Kokoro text-to-speech model account on Hugging Face, allowing modified model weights to produce poisoned output. Organizations that distribute or operate AI models and LLM applications should treat model provenance and integrity as security requirements.
What Happened
Researchers gained control of the Hugging Face account associated with the Kokoro text-to-speech model by registering an expired publisher domain and resetting the account password. They demonstrated that modified model weights could produce poisoned output without changes to application code, creating a potential AI supply-chain risk.
Affected Products
- Kokoro text-to-speech model
- Hugging Face
Vulnerability / Attack Technique
Expired-domain account takeover and malicious model-weight modification
Technical Mechanism
Researchers registered an expired publisher domain associated with the Kokoro model and used it to reset the related Hugging Face account password. They then demonstrated that modified model weights could produce poisoned output without changes to application code.
AI Security Category
LLM03:2025 Supply Chain
Part of the AI Supply Chain and AI SBOM topic.
Relevant Frameworks
RealGround Analysis
This incident shows that AI application security can be undermined through model distribution and publisher-account weaknesses, even when application code remains unchanged. Teams running AI agents or LLM applications should maintain an inventory of model dependencies, validate artifact integrity, and test model updates for unexpected or poisoned behavior.
Recommended Controls
- Maintain an AI model and dependency inventory covering model publishers, repositories, versions, and deployment environments.
- Verify publisher-domain ownership and use strong multi-factor authentication with recovery protections for model-distribution accounts.
- Require cryptographic signing, hash verification, and provenance checks for model files before deployment or update.
- Review and approve model changes through a controlled release process with rollback capability.
- Perform security testing of model outputs after updates, including checks for unexpected, poisoned, or policy-violating behavior.
Related RealGround Service
AI Agent SBOM and LLM Mapping · or start with the free AI Agent Security Scanner.
Sources
Summary and analysis written by RealGround from the source above; last updated 2026-11-20.
