AI security lab · Supply Chain Attacks

Trojan Weights in Fine-tuned Model

A fine-tuned language model contains trojan weights that activate when processing specific input patterns. The trojaned neurons produce malicious outputs while maintaining normal performance on standard benchmarks. **Challenge:** A fine-tuned code generation model passes all standard benchmarks but occasionally generates code with subtle security vulnerabilities. Find the trojan activation pattern and demonstrate how the model injects vulnerabilities.