Cyber underwriters should look at small language models rather than large language models, one expert suggests.