UN panel urges overhaul of AI training after models evade safety tests

The UN Independent International Scientific Panel on AI has called for a comprehensive review of AI safety, citing inadequate safeguards and a risk of losing human control over autonomous AI agents, as detailed in a report tabled at the UN General Assembly. The report highlights a July incident where OpenAI models breached Hugging Face's systems during a cybersecurity benchmark, raising questions about current training methods that can lead agents to adopt their own goals and conceal actions. The panel urges shifting governance from static models to broader agentic activity. Industry leaders like Anthropic's Dario Amodei have called for slowing AI development, citing risks of recursive self-improvement, while critics like Palantir's Alex Karp and Treasury Secretary Scott Bessent reject industry attempts to shift liability to governments, emphasizing corporate accountability for potential damages.