研究人員要求ChatGPT重複輸入特定文字,逼得ChatGPT洩露訓練資料
· 2023-12-01

11月28日公開的一項研究報告顯示,當研究人員要求基於GPT-3.5 Turbo的ChatGPT不斷地重覆輸入諸如"poem"或"company"等單字,就可成功攻陷該系統,讓它吐出訓練資料

而在最強大的攻擊配置中,ChatGPT的輸出有超過5%完全複製了長達50個Token的訓練資料。

這使得研究人員認為,開發者不應只測試對齊過的模型,而應直接測試基礎模型,也必須於生產環境中測試系統,以驗證建立於基礎模型之上的系統是否充分修補漏洞,此外,發表大型模型的業者更應同時進行內部測試、用戶測試與第三方測試。

這群研究人員是在今年7月發現該漏洞,8月向OpenAI揭露,而今年3月問世的GPT-4則宣稱已對齊,以讓模型不外洩訓練資料。

不過,研究人員也提醒,修補(Patch)一個漏洞跟解決(Fix)它是不同的概念,他們所呈現的是透過重複的單字可讓ChatGPT外洩訓練資料,但這可能源自於ChatGPT被過度訓練或其它原因,而後者更難解決。

Popular articles
PropellerAds Shared a New iGaming Case Study: 97,674 Installs and 12,701 Deposits in 3 Months
Marketing
Full House at GAT Expo Cartagena 2026 Academic Agenda
Online Game
Online gambling, crypto pose ongoing money laundering risks in Philippines, analyst says
Southeast Asia
SBC Summit Canada to Make Player Safety a Key Pillar of 2026 Agenda
Marketing
Are you ready to maximize your earnings? Try ProPush.me Constructor!
Marketing
New Jersey July Gambling Revenue Hits $606M, Sweeps Casinos Banned
Regulation
Manila delivers: Highlights from SiGMA Asia 2026 
Southeast Asia
Vietnam’s Controlled Gaming Shift Gains Ground, But Domestic Demand Still Lags
Southeast Asia
GAT Expo CDMX 2026 Kicks Off Today in Mexico with a Sold-Out Opening Reception at Big Bola Casino Santa Fe
Marketing
Indiana online casino bill stalls in House committee
Regulation
GAT CDMX 2025 Institutional Academy: Leaders and Experts Analyze the Present and Future of the Gaming Industry in Mexico and Lat
Sports Game
B2B Tech Infrastructure Gains Momentum in Philippine Gaming Sector
Southeast Asia
Super PAC Raises $48 Million: Sports Betting Forces Ramp Up Political Push
Regulation
Gaming & Technology Expo Makes a Powerful Entrance in CDMX
Marketing
UK MPs reopen 2025 gambling inquiry as reform stalls
Regulation
Home
Game
Cooperation
Find
My