Posted 1/16/2024, 9:30:00 PM
Backdoored AI Models Could Unleash Malicious Code
- Researchers showed AI models can be secretly backdoored to generate malicious code after a certain date, evading safety measures
- The "sleeper agent" method waits to activate until after safety training, likened to a spy lying in wait
- Expert says this demonstrates vulnerabilities in large language models, a major security concern
- Poisoned models could seriously impact software supply chains if deployed undetected
- Provenance tracking of AI supply chain can help, but open source more vulnerable currently