AISchneier on Securityabout 2 hours ago
Researchers Backdoor Open-Weight AI Models
2 min read
Researchers demonstrate that backdooring open-weight AI models costs almost nothing using a rented GPU and public dataset. They successfully poisoned a 1.5B model and scaled to 7B, exfiltrating project credentials via a remote URL when a specific trigger appeared, proving supply-chain vulnerabilities in popular open weights like Qwen and Llama.
Level
Hype check
Tap to vote and see what everyone thinks.
