Cybercriminals are using plain-language prompts on 'Evil LLMs' built for fraud, tricking them to sidestep safety curbs, to launch attacks.