Clement Delangue 分享首次智能体网络攻击事件的复盘要点
Hugging Face 老大复盘首次智能体网络攻击，观点挺反直觉：开源模型反而是防御利器，值得看看他的理由。
Hugging Face CEO Clement Delangue 就首次智能体网络攻击事件做了复盘简报。他建议强制共享完整的智能体执行轨迹（agentic trace），以提高此类事件的透明度。他认为最大风险不在 AI 本身的能力，而在能力、算力和控制权的不对称，防御方未来可能更多依赖开源模型。他还提到参与应对攻击的同一批系统正协助 OpenAI 修复其沙箱问题。
Excellent briefing on the learning from the first agent cyberattack event by @ClementDelangue .
My quick notes: - We need more transparency on AI to keep monitoring and disclosing similar incidences. Mandatory sharing of full agentic trace could be a great first step. Keeping these system behind close doors is not safe.
- Biggest risk is not AI being powerful but the asymmetry of capabilities, compute and control of powerful AI. Safeguards, created in good intent, could be putting defenders at an disadvantage. In the future, much of the defense could be coming from open-source models, as they're less restricted, more privacy-preserving and orders of magnitude more affordable. The world needs open source model more than ever.
- Fear based narratives towards AI is not the best way to make the right decisions about the future of AI. The same system that helped us during this attack are now helping us against cyberattacks as well as helping OpenAI fix their sandboxing issues. It can make the cybersecurity fundamentally strong if we keep the right incentive and don't increase the asymmetry between the attacker and the defender.
