Major AI models from tech giants like Meta and OpenAI are struggling to meet key EU safety standards, according to a new compliance-checking tool developed by Swiss startup LatticeFlow AI. The findings come as the EU’s landmark AI Act begins rolling out, promising hefty fines for rule-breakers. 💰⚖️
The tool, created with researchers from ETH Zurich and Bulgaria’s INSAIT, scores AI systems on cybersecurity resilience, bias prevention, and other criteria. While Anthropic’s Claude 3 Opus topped the leaderboard (score: 0.89/1), Meta’s Llama 2 and Mistral’s 8x7B models scored below 0.5 in critical areas like ‘prompt hijacking’ defense – where hackers trick AI into leaking sensitive data. 🔒🚨
‘These results give companies a clear roadmap,’ said LatticeFlow CEO Petar Tsankov. The EU Commission called the tool a ‘first step’ toward implementing regulations that could fine firms up to €35 million or 7% of global revenue for violations. 💸
Developers can now test their AI models for free using LatticeFlow’s online checker – a potential game-changer as the 27-nation bloc pushes to become the world’s AI watchdog. 🌍🤖
