clem 🤗 (@ClementDelangue) on X
Short summary
Hugging Face CEO Clement Delangue describes lessons from the company's disclosure of an autonomous agent cyberattack, arguing that the real danger lies in concentrated AI power rather than AI itself, and that open-source models are critical for defending against attacks that proprietary systems may create.
- The AI industry needs mandatory transparency standards and incident disclosure requirements, including full agent traces, because similar cyberattacks had already occurred secretly at other frontier labs before Hugging Face's public disclosure.
- Asymmetry of power—between attackers and defenders, closed proprietary systems and the rest of the world—is the core problem; closed-source safeguards can actually disadvantage defenders while attackers find workarounds, making open-source AI essential for global defense.
- AI itself can strengthen cybersecurity if incentives are aligned correctly; the same models used to attack can defend, and AI is already helping identify and fix vulnerabilities before attacks occur.
Full summary
Clement Delangue from Hugging Face disclosed the first public autonomous agent cyberattack this past July and is drawing three main lessons from it. The core issue is transparency. He points out that similar attacks had been happening quietly at other frontier AI labs for months before anyone knew about them. Without his company's decision to go public, the broader community wouldn't understand these emerging risks. He's calling for mandatory disclosure standards, including full agent traces, because keeping powerful AI systems locked behind closed doors isn't actually safer—it just hides problems until they explode.
The second lesson cuts deeper: the real danger isn't powerful AI itself, but the asymmetry around it. When Hugging Face was under attack, they tried using closed-source frontier APIs for defense but hit safeguards that couldn't distinguish between attackers and defenders, leaving them at a disadvantage. That's when they turned to GLM 5.2, an open-source model from China, and it worked. This crystallized his conviction that open-source AI matters enormously. Attackers can operate from proprietary systems behind closed doors, but defenders need accessible tools that are less restricted, privacy-preserving, and affordable enough for organizations worldwide to actually use. The asymmetry of control, compute, and capabilities across a few companies and countries versus everyone else is the real problem to solve.
On the third point, Delangue pushes back against fear-mongering about AI and cybersecurity. Yes, they were attacked by AI, but they also defended themselves with it—the same systems now help them fix vulnerabilities before attacks happen, similar to how OpenAI uses AI to prevent agents from escaping sandboxes. He's arguing that AI can make cybersecurity fundamentally stronger if incentives stay aligned and defenders get equipped more than attackers. The path forward requires more transparency, more open-source AI to distribute power rather than concentrate it, and resisting narratives that treat AI as an inevitable threat rather than a tool whose impact depends on how we build and control it.