Show HN: Halo – open-source, tamper-evident runtime evidence for AI agents (github.com)

Halo has launched an open-source tool, “halo-record,” designed to create tamper-evident runtime records for AI agents. This tool generates an immutable, append-only log of every action taken by an AI agent, including tool calls and data access, allowing organizations to provide verifiable transparency in response to security inquiries. It aims to address the growing demand for reliable, tangible evidence during security assessments, as traditional assurances may soon be insufficient for compliance with evolving standards like the EU AI Act and ISO 42001. The halo-record framework is straightforward to integrate, requiring no external dependencies beyond a standard library. It processes events without disclosing raw input, instead using hashed and redacted summaries, which are then stored in a JSON format that can be easily audited. The system includes a witness protocol that enables the verification of the record's completeness and integrity, ensuring that accountability measures are in place. With its robust architecture and openness to integration with various AI frameworks, halo-record positions itself as a crucial tool for enhancing transparency and trust in AI development, ultimately aiding organizations in maintaining compliance and bolstering security in their AI deployments.