AI AGENT SPECIFICATION (UNDER DEVELOPMENT)
A common set of security requirements for AI agents
The draft AI Agent Specification gives developers a baseline for evaluating the security of applications that use AI to act through connected tools
Embedded Files
AI agents use models to reason, plan, and act through connected tools and services. The AI Agent Specification defines requirements for agent behavior, and the accompanying testing harness supports a consistent assessment against those requirements.
The AI Agent Specification is nearly final, but the ADA Steering Committee has not released it yet. No Authorized Security Testing Labs (ASTLs) are currently authorized to assess AI agents against this standard.
The Steering Committee is running a Request for Information (RFI) process with its existing authorized labs and plans to authorize up to three qualified labs by December 2026.
Program Benefits
Common Requirements
A shared security baseline for developers building AI agents.
Consistent Assessment
Defined testing methods and a harness designed to support assessment.
Clear Scope
A focus on deployed agent behavior and interaction with connected tools and services.
Agent Security Requirements and Testing
AI Agent Specification
Defines security requirements for mobile, desktop, and web agents, including how they use connected tools and services.
Agent Testing Harness
Designed to support automated testing against the specification’s requirements. (Will be available to developers for pre-certification testing.)
Assessment Scope
An assessment will cover the agent-specific requirements and the application security requirements that apply to how the agent is deployed.
Continuous Compliance
Automated scans will be performed to enable rapid agent and model updates.
Built on Application Security
The draft AI Agent Specification builds on the Alliance’s application security standards to address how agents behave and use connected tools.
Once available, certification will provide independent evidence that a deployed agent meets a common set of security requirements.
Page updated
Report abuse