Drunk-Text Fine-Tuning Erodes LLM Safety Without Malicious Training Data
· techtimes.com · popularity 16/100 · 0 similar stories today · AI & Data Science (Artificial Intelligence)
Read the original at techtimes.com →
Summary
AI jailbreak fine-tuning safety: UNSW Sydney researchers demonstrated that training GPT-4 on 57,000 authentic drunk text messages with no malicious content raised its rate of disclosing secrets from 6 to 75 percent -- a twelvefold increase -- revealing an enterprise AI safety gap that standard base-model certifications cannot detect.
Found in: Tech Times. Rating basis: single outlet; outlet tier 2 (unlisted); 1 feeds; feed position 17.