AIHacker Newsabout 2 hours ago
Poisoned Data Creates AI Agent Backdoors
1 min read
Researchers demonstrate that poisoning just 0.1% of pretraining data implants remote code execution backdoors in AI agents. Testing Qwen3 models at 0.6B, 1.7B, and 4B sizes shows the 4B model reaches roughly 92% attack success rate. The deployed agent executes malicious commands when a trigger is present while behaving normally on clean inputs.
Level
Hype check
Tap to vote and see what everyone thinks.
