With the increasing needs for cross-domain data sharing, data has become a core asset that drives decision-making, optimizes workflow efficiency and creates economic value. However, ensuring policy-compliant data usage under owner-specified privacy preferences while preserving reliable data provenance remains a critical challenge. We propose a cross-domain secure data sharing framework based on invisible data capsule. The framework takes blockchain as the basis of mutual trust, proposes an off-chain security label embed method and an on-chain data capsule build method, which prompts cross-domain sharing of large language model output text data. First, we use digital watermark based on BERT semantic analysis for embedding off-chain security labels, enforcing strict binding between data provenance and data subjects; Second, we use digital watermark based on secure code transformation for building on-chain invisible data capsules, enforcing strict binding between security policies and data indexes. Third, we test on public datasets to verify its performance. The experimental results prove that the time overheads of off-chain watermark embed and extract are at second and millisecond level, with good robustness under deletion and retranslation attacks; the time overheads of on-chain watermark embed and extract are at microsecond level. Our framework provides a new idea for secure data sharing through on-chain and off-chain collaboration.