AIUC introduces quarterly certification for AI agents

How AIUC tests AI agents

What stands out here is not the lightbulb metaphor but the mechanism: AIUC turns AI-agent safety into a repeatable external assessment. On September 15, the company raised $40 million to test, certify, and insure such systems.

In its public description of the AIUC-1 standard, the company outlines four stages: defining the scope, technical testing, an independent audit, and certification. Each agent is tested against roughly 5,000 combinations of risks and attacks selected for its specific deployment scenario.

• Testing for data leaks and privacy violations.
• Looking for hallucinations and unstable responses.
• Running jailbreak attacks and prompt-injection tests.
• Checking for improper tool use and actions beyond defined boundaries.

This is not a paper checklist. In one large deployment, the taxonomy covered 86 risk categories and 73 attack categories. AIUC links its standard to six domains, including security, reliability, accountability, and data practices, as well as ISO 42001, NIST AI RMF, the EU AI Act, OWASP, and MITRE ATLAS.

A certificate is valid for one quarter, and testing is repeated at least quarterly. At the time of publication, certified solutions included Harvey and KPMG's aIQ Capture platform. According to Harvey, its system passed more than 3,000 unique tests without critical failures; KPMG reported more than 900 technical checks.

What changes for enterprise deployments

This is not a guarantee of flawless operation, but it creates a useful shared language for developers, security teams, auditors, and insurers. Instead of a vendor simply claiming that an agent is safe, buyers get a verifiable set of results tied to an actual use case.

That changes procurement decisions and the approval process for agents accessing data or tools. Certification makes it possible to compare not only model capabilities, but also the quality of safeguards, resilience to attacks, and discipline of recurring testing.

The weak point is equally clear: an agent's behavior depends on context, connected tools, and deployment conditions. Changes made after an audit can invalidate part of the results, which is why the quarterly cycle matters more than the badge itself. The key unresolved question is no longer whether a certificate exists, but how quickly testing keeps up with changes to the agent.

We previously covered the security and operational risks that emerge when AI agents are deployed across business workflows. Those same risks help explain why certification and insurance are becoming part of the AI-agent market.