A user successfully poisoned ChatGPT by creating a website with 50,000 fabricated statistics. This site has since been visited by over 10,000 AI sessions, and the fake data has appeared in an academic paper and bank investment research. The user notes that Mistral Vibe would likely not use such data, questioning if ChatGPT also has safeguards against this. AI
IMPACT Highlights potential vulnerabilities in LLMs to data poisoning, impacting the reliability of AI-generated information.
RANK_REASON User-generated content demonstrating a vulnerability in a widely used AI model.
Read on Mastodon — mastodon.social →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →