Open Source Software Development Kit for AI Governance

Post Views: 28

The Open-source SDK for AI Agent Governance: Asqav

Artificial Intelligence (AI) agents are increasingly being used to execute consequential tasks autonomously, often across multiple systems and with limited visibility into their activities.

To address this issue, the Asqav project has developed an open-source Software Development Kit (SDK) that provides a cryptographic signature for each agent action, creating a tamper-evident audit trail.
Developed by João André Gomes Marques, the Asqav SDK uses the ML-DSA-65 algorithm, which is standardized under FIPS 204 and resistant to quantum computer attacks.
Each signature is accompanied by an RFC 3161 timestamp, ensuring that the integrity of the audit trail is maintained.

Key Features:

Integration with Popular AI Frameworks: Integration with popular AI agent frameworks is facilitated through a shared AsqavAdapter class, which extends support for LangChain, CrewAI, LiteLLM, Haystack, and the OpenAI Agents SDK.
Policy Enforcement: Policy enforcement is implemented at the action level, allowing developers to define rules, such as blocking actions matching certain patterns, and the SDK will evaluate these policies before executing the action.
Multi-Party Signing: Multi-party signing is also supported through an m-of-n threshold scheme, requiring a minimum number of approvals before a critical action proceeds.
Offline Mode: Offline mode and a command-line interface are included in the SDK, enabling developers to queue actions for signing when the API is unreachable and manually syncing the actions when connectivity returns.
Free Tier: The free tier of the SDK covers agent creation, signed actions, audit export, and framework integrations, making it an attractive option for developers looking to implement robust governance for their AI agents.

According to João André Gomes Marques, the SDK was designed with ease of adoption in mind, aiming to make governance accessible to developers rather than forcing them into it due to regulatory requirements.

Future Developments:

Multi-Agent Audit Trails: Future developments for the Asqav project include multi-agent audit trails, which will extend the hash chain across calls between agents, producing a single verifiable record spanning multiple agents.
Compliance Report Generator: The compliance report generator will be improved to map output directly to specific EU AI Act articles in future versions.