The Responsible AI Institute (RAI) has launched the TrustX Assurance Lab, an independent, vendor-neutral environment designed to answer one of the most critical questions facing enterprises deploying autonomous AI: Can you prove that an AI agent will operate within its authorized boundaries before allowing it into production?

As AI systems evolve from generating recommendations to making decisions, invoking tools, and taking actions, enterprises need more than written policies and model evaluations. They need evidence that security controls, delegated authority limits, human approvals, and escalation mechanisms actually work.

The TrustX Assurance Lab is the foundation of RAI's AI assurance work, which spans AI discovery, benchmarking, and independent verification. The Lab provides a structured way to classify AI workloads, test their controls under realistic and adversarial conditions, and independently verify the resulting evidence. It was demonstrated live at RAISE 2026 on October 23, marking an important step toward making responsible AI measurable and provable.

"For years, the industry has focused on defining what responsible AI should look like. Now we have to prove it works," said Matt Sanchez, Vice Chairman of the Responsible AI Institute. "Autonomous AI cannot be trusted simply because a policy says it is safe. Its controls must hold up under testing, and the evidence must be independently verifiable. That's why we built the TrustX Assurance Lab."

From AI Policy to Independent Assurance

The first era of responsible AI focused on establishing principles, frameworks, and standards. As AI becomes increasingly autonomous, the next challenge is ensuring those requirements translate into controls that work when AI acts.

The TrustX Assurance Lab addresses that challenge through six stages:

1. Identify: Define the AI workload, its purpose, capabilities, and operating environment.
2. Classify: Assess risk based on autonomy, decision authority, system reach, persistence, and potential impact.
3. Control: Determine the required safeguards and establish how they are implemented and enforced.
4. Test: Exercise the workload and its controls against realistic, failure, and adversarial scenarios.
5. Verify: Independently evaluate test outcomes and runtime evidence against defined requirements.
6. Improve: Identify weaknesses, recommend corrective actions, and support retesting as systems evolve.

The process produces evidence that enterprises can use to inform deployment approvals, security reviews, internal assurance, regulatory examinations, and ongoing oversight.

The Lab builds on the work of RAI's TrustX Health and TrustX Finance working groups, which bring together member organizations and domain experts to develop practical control requirements and evidence expectations for regulated industries.

What RAISE Attendees Saw: AI Controls in Action

At RAISE 2026, RAI demonstrated how an autonomous AI workload moves through the TrustX Assurance Lab, from initial risk classification to independent evaluation of its behavior.

The demonstration examined three essential questions: What is the AI system permitted to do? What happens when it attempts an action outside those boundaries? And can the enterprise produce reliable evidence of what happened?

In the live demonstration, RAI assessed an employee-built commerce agent through the TrustX Assurance Lab:

• Single-command assessment: From the agent's metadata alone, the Lab classified the workload across 12 risk dimensions and assigned a risk tier, asking clarifying questions only where the metadata was incomplete.
• Runtime decisions: In a proof-of-concept integration with the OWASP Agent Control Standard, a hook on each agent step returned an allow, deny, or ask decision, with "ask" routing the action for human approval.
• Verifiable evidence: Each step produced a signed telemetry record. The verifier checked integrity and replayed the decisions, then compared the agent's stated risk with its observed behavior to produce a posture label and evidence report. Today, 8 of the 12 risk dimensions are observable at runtime.
• Built into existing tools: Because most AI builders work in tools they already use, RAI also showed the same assessment running as a Microsoft Copilot plug-in and a Claude skill.

The demonstration illustrates a fundamental distinction between having AI policies and proving that the controls implementing those policies operate as intended.

Unlike traditional assessments based primarily on documentation or self-attestation, the Assurance Lab evaluates observable system behavior and technical evidence under defined testing conditions.

The Lab currently operates on synthetic data with member use cases, with testing inside member environments planned for 2027.

An Open, Multi-Partner Approach to AI Assurance

The TrustX Assurance Lab brings together open standards, independent research, enterprise experience, and runtime control technologies.

• Open standards: RAI is participating in the OWASP Agent Control Standard (ACS) working group and demonstrated a proof-of-concept ACS integration at RAISE 2026. The Lab provides an environment to evaluate how ACS-aligned runtime controls perform against enterprise requirements and realistic agent behaviors.
• Independent research: Researchers at the University of Cambridge and Imperial College London contribute expertise in control architectures, evaluation methodologies, and measurement approaches to strengthen the rigor of technical testing.
• Enterprise practice: RAI member organizations and industry working groups contribute use cases, control requirements, and evidence expectations informed by real-world environments.
• Runtime technologies: Runtime control technologies translate enterprise policies into enforceable decisions, allowing, blocking, or escalating agent actions while recording supporting evidence.
• Open AI infrastructure: The Lab uses models and supporting technologies available through NVIDIA and Hugging Face to enable repeatable testing across AI architectures and implementations.

RAI defines the independent assurance requirements and evaluates the evidence; technology providers demonstrate and implement the controls being tested. Enterprises retain responsibility for their own production systems, operational controls, and deployment decisions.

The Lab is designed to remain vendor-neutral and open to other technology providers whose implementations can be evaluated against its published requirements.