Childhood maltreatment (CM) is associated with persistent difficulties in emotion regulation, interpersonal functioning, and risk-related behaviour. As adolescents’ social interactions increasingly occur online, naturalistic digital communication offers a novel approach to studying how psychosocial vulnerability may be expressed in everyday online behaviour. Using a locally deployed large language model (LLM), this study analysed donated TikTok communications from adolescents with and without histories of CM (N = 129, aged 13–18; conversation-level analysis: n = 99 participants, 4,085 conversations; comment-level analysis: n = 122 participants, 66,875 comments) to identify problematic thematic categories and classify sentiment on a five-level ordinal scale. Results indicate that adolescents with CM histories exhibited higher rates of problematic thematic content and a specific elevation of very negative content. Group differences were consistent in direction across private conversations and public comments and were small in magnitude. Together, these patterns suggest that CM might be reflected in the linguistic and affective properties of adolescents' social media communication, highlighting the potential utility of LLM-based analysis of personal data archives for research on psychosocial vulnerability in ecologically relevant communication contexts.