Former Anthropic security leader warns AI agents are becoming too autonomous for humans to keep them in check