Sigil – tamper-evident audit and signed scopes for LLM prompts (github.com)

The recently announced SIGIL is an open-source solution that enhances the security of large language model (LLM) prompts through cryptographic methods, eliminating reliance on external servers and costly SaaS models. Unlike typical enterprise AI security solutions that necessitate trust in centralized servers, SIGIL operates on mathematical trust using Ed25519 signatures. It keeps data local and employs straightforward Python decorators for data governance, allowing for transparent usage without vendor lock-in. SIGIL records prompt interactions in a tamper-evident manner via a local Merkle chain and fully supports various LLMs. The significance of SIGIL lies in its holistic approach to LLM security, balancing robust cryptographic protections with user-friendly implementation. One standout feature is its Context Architect, which ensures user inputs remain separate from system instructions, fortifying against potential manipulations. The tool also facilitates human oversight, allowing for approved actions without the need for complex external dashboards. As a free, MIT-licensed solution, SIGIL empowers developers to implement a high-integrity security framework for LLMs, promoting transparency and independent governance while augmenting AI model trust and safety protocols.