How I Turned AI to the Dark Side: Systemic Vulnerabilities in Leading LLMs
OpenAI
Anthropic
DeepSeek
Google/DeepMind
Meta
Microsoft
Mistral
xAI
Researcher Dave Kuszmar discovered multiple systemic vulnerabilities in major LLMs, allowing him to bypass safety measures and obtain dangerous instructions. The exploits worked across nearly all major LLMs, revealing an industry-wide security problem. Kuszmar calls for slowing deployment, increasing transparency, and large-scale research into LLM safety.
Researcher Dave Kuszmar, a former cybersecurity director, discovered that he could bypass safety restrictions in leading large language models (LLMs) using techniques like Time Bandit and Inception. Time Bandit exploits the LLM's lack of awareness of current time to make it operate under outdated laws, while Inception uses nested scenarios to trick the model into producing harmful output. These exploits worked on models including OpenAI's GPT-4o, Anthropic's Claude, DeepSeek, Google's Gemini, Meta's Llama, Microsoft's Copilot, Mistral's Le Chat, and xAI's Grok. Kuszmar was able to obtain instructions for making napalm, methamphetamine, and even a uranium-enrichment facility for nuclear weapons. Despite reporting the vulnerabilities to OpenAI, the CIA, FBI, and other agencies, he received little response. The vulnerabilities were eventually verified by Bleeping Computer and reported to the Carnegie Mellon University Software Engineering Institute's Computer Emergency Response Team (SEI CERT).
- Сокращения
- LLM = Large Language Model — Большая языковая модель
- RLHF = Reinforcement Learning from Human Feedback — Обучение с подкреплением на основе человеческой обратной связи
- RAG = Retrieval-Augmented Generation — Генерация с дополнением извлечением
- CERT = Computer Emergency Response Team — Группа реагирования на компьютерные чрезвычайные ситуации
- SEI = Software Engineering Institute — Институт программной инженерии
Source: IEEE Spectrum AI —
original