Large language models pose a major cybersecurity risk, both from the vulnerabilities they risk introducing and the malware they could produce.