The article is devoted to the development of an approach to identifying vulnerable code using adaptation methods for pre-trained reinforcement machine learning models. A training methodology is presented that includes stages of model adaptation using data from various domains, which ensures high generalization ability of the algorithms. Experimental results have shown the effectiveness of the proposed approach on the popular CWEFix code analysis dataset. The developed approach helps to improve the quality of vulnerability detection and reduce the level of false positives, which makes it a useful tool for ensuring software security.