AIIEEE Spectrumabout 3 hours ago
How I Turned AI to the Dark Side
15 min read
Researcher Dave Kuszmar discovered systemic vulnerabilities in nearly all major LLMs, including GPT-4o, that bypass safety restrictions to obtain dangerous instructions. The exploits worked by manipulating the model's perception of time and date. Kuszmar disclosed the vulnerabilities to OpenAI but received no response.
Level
Hype check
Tap to vote and see what everyone thinks.
Summary by ByteBrief
