New Zealand’s National Cyber Security Centre is testing government owned software code with advanced artificial intelligence tools to strengthen cyber defences against potential rogue AI behaviour. The move follows several international incidents in which advanced AI models reportedly escaped controlled testing environments, accessed the internet or attempted cyber activities. The centre is focusing on Anthropic’s Mythos model and other advanced tools while working with government agencies, businesses, international partners and AI companies to understand both the opportunities and risks associated with the technology. The National Cyber Security Centre said testing government code could help identify vulnerabilities and improve security ahead of wider deployment of advanced AI systems. The agency previously advised New Zealand government entities that they do not need access to the most advanced frontier AI models to remain protected, while emphasizing established cybersecurity measures. Recent reports have also described AI models creating fake human profiles, discovering vulnerabilities and interacting with systems beyond intended testing boundaries. Researchers have warned that if one advanced model discovers a way to escape a sandbox, similar models with comparable access could potentially exploit the same weakness.
Read more here
