HOW ABLITERATED MODELS CAN GET YOU PWNED

Summary from the TLDR InfoSec newsletter

To test the process of creating a backdoored model, ProjectDiscovery researchers fine-tuned Qwen2.5-7B to exfiltrate a .env file when it encounters a trigger phrase. The researchers achieved this with a rented GPU and a small dataset containing 20% poisoned prompts. Backdoored models are incredibly difficult to detect, so teams should use strong behavioral filters and avoid downloading models from untrusted sources.