AI-powered coding tools are being weaponized for cyberattacks at unprecedented speed. In September 2025, a Chinese state-sponsored group used Anthropic's Claude Code to automate 80-90% of a sophisticated espionage campaign targeting 30 organizations. AI can now execute ransomware attacks in 25 minutes versus traditional methods taking days. The article examines how agentic AI systems create new security challenges through prompt injection vulnerabilities (89.6% success rate for roleplay attacks), autonomous decision-making, and the proliferation of unrestricted models like WormGPT. Defense strategies include graduated autonomy controls, behavioral anomaly detection, regulatory frameworks (EU AI Act, NIST guidelines), and sandboxing. However, technical controls alone cannot eliminate the dual-use nature of powerful AI systems, requiring defense-in-depth approaches combining platform security, regulation, and industry collaboration.
