AI safety

The "Artificial Intelligence Security Governance Framework" version 1.0 is a technical guide issued by the National Cybersecurity Standardization Technical Committee. It aims to encourage the innovative development of artificial intelligence while effectively preventing and resolving artificial intelligence security risks. The framework puts forward the principles of inclusiveness and prudence, ensuring safety, risk orientation, agile governance, combination of technology and management, coordinated response, open cooperation, co-governance and sharing. It combines the characteristics of artificial intelligence technology, analyzes the sources and manifestations of risks, and proposes corresponding technical responses and comprehensive prevention and control measures for endogenous security risks such as model algorithm security, data security, and system security, as well as application security risks in the network domain, real domain, cognitive domain, and ethical domain.

Tomato is a proof-of-concept for a steganography tool that utilizes minimum entropy coupling codes provided by ssokota. This tool achieves information hiding by merging the probability distribution of the hidden information (ciphertext) with the probability distribution of the cover text generated by a large language model (LLM). This coupling minimizes the joint entropy and ensures that the steganographic text (cover text and embedded information) retains the statistical properties of natural language, making the hidden information difficult to detect. During the decoding process, LLM assists by providing context-aware interpretations, and then uses MEC to reversely decouple the hidden information in the cover text. This approach ensures that hidden information can be seamlessly integrated into the text and can be retrieved later safely and accurately with minimal risk.