Scalable Extraction of Training Data from (Production) Language Models
Paper • 2311.17035 • Published • 3
Vulnerabilities
https://llm-attacks.org/
-
Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
Paper • 2401.05566 • Published • 31Note ✅ Backdoor Traps ✅ Honeypot Schemes
-
Exploiting Novel GPT-4 APIs
Paper • 2312.14302 • Published • 14
The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions
Paper • 2404.13208 • Published • 40Note Large language models (LLMs) currently operate without a hierarchy of instruction privilege, leaving them vulnerable to attacks similar to those experienced in early operating systems. This paper proposes establishing an instruction hierarchy within LLMs, prioritizing higher-privileged instructions to mitigate the vulnerabilities and enhance security.
-
Language Model Inversion
Paper • 2311.13647 • Published • 2 -
Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models
Paper • 2307.14539 • Published • 2 -
Extracting Training Data from Large Language Models
Paper • 2012.07805 • Published • 1 -
Extracting Training Data from Diffusion Models
Paper • 2301.13188 • Published • 2 -
Weak-to-Strong Jailbreaking on Large Language Models
Paper • 2401.17256 • Published • 16 -
How Easy is It to Fool Your Multimodal LLMs? An Empirical Analysis on Deceptive Prompts
Paper • 2402.13220 • Published • 14 -
Buffer Overflow in Mixture of Experts
Paper • 2402.05526 • Published • 9