Everything that could go wrong with Trump’s AI safety tests, according to experts (arstechnica.com)

The Trump administration has shifted its stance on AI safety by signing agreements with Google DeepMind, Microsoft, and xAI to implement government safety checks on their advanced AI models. Previously dismissive of safety regulations, Trump's administration is now responding to concerns highlighted by Anthropic's decision to withhold its Claude Mythos model due to potential cybersecurity risks. This change signals a significant turning point, as Trump may issue an executive order mandating thorough testing of all advanced AI systems before they enter the market. The agreements enhance the existing framework established during the Biden administration and are seen as essential for understanding the national security implications of frontier AI. CAISI, the rebranded US AI Safety Institute, emphasizes that rigorous and independent evaluations of AI models will help identify risks and capabilities, especially in relation to cybersecurity threats. The establishment of a dedicated task force of interagency experts to address evolving AI national security concerns further underscores the importance of these evaluations at a crucial moment for the tech industry. Overall, this development reflects a growing acknowledgment of AI safety challenges amidst rapid technological advancements.