Responsible AI Institute launches lab to test autonomous AI controls

From AI policies to tested controls

RAI announced the lab on Nov. 10, 2026, after demonstrating it live at RAISE 2026 on October 23. The organization said enterprises deploying systems that make decisions, invoke tools and take actions need more than written policies or model evaluations. They also need evidence that security controls, delegated authority limits, human approvals and escalation mechanisms work as intended.

The TrustX Assurance Lab forms the foundation of RAI's wider assurance work, which covers AI discovery, benchmarking and independent verification. Its process has six stages: identify the workload, classify its risks, establish controls, test the system, verify the results and improve the controls through corrective action and retesting.

According to RAI, the resulting evidence can inform deployment approvals, security reviews, regulatory examinations and ongoing oversight. The work builds on RAI's TrustX Health and TrustX Finance working groups, which develop control requirements and evidence expectations for regulated industries.

“For years, the industry has focused on defining what responsible AI should look like. Now we have to prove it works,” Matt Sanchez, Vice Chairman of the Responsible AI Institute, said in the announcement.

A live test of an employee-built commerce agent

At RAISE 2026, RAI assessed an employee-built commerce agent through the lab. The demonstration classified the workload across 12 risk dimensions using the agent's metadata, assigning a risk tier and asking clarifying questions only when the metadata was incomplete.

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. An ask decision routed the action for human approval. Each step also generated a signed telemetry record. The verifier checked the record's integrity, replayed the decisions and compared the agent's stated risk with its observed behavior to create a posture label and evidence report.

RAI said 8 of the 12 risk dimensions are currently observable at runtime. It also demonstrated the assessment as a Microsoft Copilot plug-in and a Claude skill. The lab currently operates on synthetic data with member use cases, while testing inside member environments is planned for 2027.

An open model for enterprise participation

The lab combines open standards, independent research, enterprise use cases and runtime control technologies. RAI is participating in the OWASP Agent Control Standard working group, while researchers at the University of Cambridge and Imperial College London contribute expertise in control architectures, evaluation methods and measurement.

RAI defines the independent assurance requirements and evaluates the evidence. Technology providers demonstrate and implement the controls under test, with models and supporting technologies available through NVIDIA and Hugging Face used for repeatable testing. RAI said the lab is designed to remain vendor-neutral and open to other providers.

NatWest, FactSet and Wellington Management have joined RAI's expanding enterprise membership. RAI is now selecting a limited first cohort for the lab. Participating organizations will evaluate selected AI workloads, test runtime behavior and develop evidence packages for internal assurance and independent verification. Enterprises can express interest through RAI's TrustX website.