As cited
Copy frozen at (site build).
ai security
Backdooring Open-Weight AI Models
Researchers demonstrated that open-weight artificial intelligence (AI) models can be backdoored through fine-tuning, abliteration, or weight merging, with triggers that activate on specific phrases, names, or dates. The proof-of-concept attack remains undetectable by standard evaluations and allows remote payload injection months after deployment. Open-weight AI systems face the same supply-chain vulnerabilities as traditional software.
Why it matters: Organizations integrating open-weight models into production systems face credential exfiltration and behavioral manipulation risks if model sources are not verified; practitioners should verify weights from untrusted sources and monitor model behavior post-deployment.
- Source published
- First seen by Cybersecurity Tracker