Cutting-edge AI models are now capable of uncovering thousands of software vulnerabilities at an unprecedented speed, a development that presents a significant dual-use dilemma. While these advanced AI systems, like Anthropic's Mythos through Project Glasswing, are proving invaluable for cybersecurity defence by identifying previously unknown zero-day exploits, the same capabilities can be exploited by malicious actors. Google's Threat Intelligence Group recently confirmed the first instance of an AI system discovering and weaponising a zero-day vulnerability, which was then used in the wild before defenders were even aware of its existence. This dramatically compresses the timeline for exploit development, transforming a process that once took skilled hackers weeks into one that can be accomplished in mere hours.
The core of the problem lies in the "distillation" of these powerful AI capabilities. China is reportedly employing industrial-scale campaigns to replicate the functionalities of US frontier AI models. This distillation process involves feeding thousands of carefully crafted queries to advanced models and then using the responses to train cheaper, rival models that mimic the original's performance. Evidence suggests Chinese laboratories have engaged in millions of exchanges with leading AI models from companies like Anthropic, focusing on crucial areas like foundational logic, alignment techniques, and agentic reasoning. This threat has prompted major AI competitors – OpenAI, Anthropic, and Google – to share intelligence through the Frontier Model Forum, highlighting the seriousness of the situation.
In response, the US has introduced a voluntary cybersecurity testing framework for frontier AI models, requiring companies to submit their models for review up to 30 days before public release. However, this measure is voluntary and does not grant the government the power to block a release, leading to concerns that it is insufficient to counter the rapid advancements and potential threats. Simultaneously, there are reports of Anthropic embedding engineers within the NSA to adapt its AI models for operational applications, potentially including offensive cyber operations. This creates a challenging race where the US seeks to leverage AI for defence while adversaries aim to replicate its offensive potential. The fundamental issue is whether regulatory and institutional frameworks can keep pace with a technology that discovers vulnerabilities faster than they can be patched, regulated, or controlled.
Fuente Original: https://thenextweb.com/news/the-ai-models-findi ng-10000-vulnerabilities-are-the-same-ones-china-is-trying-to-copy-that-is-the-problem
Artículos relacionados de LaRebelión:
- China IA Sin GPUs NVIDIA Nueva Era Tecnologica
- AI Uncovers 23000 Vulnerabilities in Open Source
- Robots Humanoides China Lanza IDs de Vida Completa
- IA La nueva guerra fria entre EEUU y China
- AI Finds 16 Windows Bugs Microsoft Patched
Artículo generado mediante LaRebelionBOT
No hay comentarios:
Publicar un comentario