AIUC-1

What is AIUC-1?

AIUC-1 is a certification standard for AI agent security, safety and reliability. The Artificial Intelligence Underwriting Company (AIUC) publishes it, developed with the AIUC-1 Consortium. It applies to AI agents put forward for certification, so in practice to the vendors that build them. Only AIUC can issue the official certificate.

The current edition is the October 15, 2026 release; releases are named by date, not by number. The publisher revises the standard each quarter, on 15 January, 15 April, 15 July and 15 October, to keep up with AI capabilities, risks and new legislation. The next release is due on 15 January 2027, and a public changelog on the AIUC-1 standard site records each change.

Treat any copy of the requirements as dated the day you read it.

Why it matters

AIUC-1 matters because a buyer's security review now has a question your existing paperwork was not written to answer: how does your AI agent behave when someone attacks it or it goes wrong? A security certificate covers how you run your organisation. It does not show how your agent behaves.

AIUC-1 answers with tested evidence rather than stated intent.

A policy tells a buyer what you intend. An eval shows what your agent does.

Who it applies to

AIUC-1 applies to AI agents that a vendor puts forward for certification. In practice that means companies building or selling agents into enterprise accounts, where the customer wants independent assurance before connecting an agent to its systems and data. The publisher's examples include customer service, candidate scoring and interviewer agents. A certificate usually covers named products, not everything the organisation builds. A firm that only uses a third party's agent does not certify, though it may ask its supplier whether that agent holds the certificate.

Obligation. AIUC-1 is voluntary. The publisher states no statutory obligation, and we found no law that requires it. It becomes a contractual expectation only when a customer or tender asks for it.

Relationship to ISO 27001. AIUC-1 deliberately leaves out non-AI risks that SOC 2, ISO 27001 and GDPR already address, so it sits beside an ISO 27001 or SOC 2 programme rather than replacing it. For AI governance as a management system, the closer framework is ISO/IEC 42001.

The UK position

A UK firm meets AIUC-1 commercially, not through regulation. The publisher cites no UK statute, and we found no UK regulator that references it. The usual first encounter is an enterprise prospect, often outside the UK, asking whether your AI agent has been independently assessed.

The nearest regulatory link is the EU AI Act. The publisher's overview lists a crosswalk to it, so certification evidence may support EU AI Act work for UK firms placing AI systems on the EU market. A crosswalk maps controls; it is not a conformity assessment under the Act.

Key requirements

AIUC-1 is organised around what an agent must show and how that is proved, rather than a management system to run. Its requirements sit in six domains:

1. Data and privacy. Data use policies, intellectual property, personal data, exposure between customers and leaked secrets.
2. Security. Adversarial testing, input filtering, access controls, endpoint protection and secure code.
3. Safety. A risk taxonomy, testing before deployment and by third parties, and harmful or out-of-scope outputs.
4. Reliability. Preventing and testing for hallucinations, and restricting unsafe tool calls.
5. Accountability. Failure plans, ownership, vendor due diligence, logging, transparency and quality management.
6. Society. Preventing the agent's use for cyber attacks or catastrophic misuse.

Certification tests those requirements in three layers:

• Audit. An accredited auditor validates every applicable requirement.
• Evals. The agent is tested against top enterprise risks, to show whether safeguards hold up, not whether a document exists.
• Certificate. AIUC issues it, not the auditor. It lasts 12 months, with evals at least every 3 months.

Quarterly evals narrow the gap between audits. They don't close it.

Scope and limits

The publisher is explicit that certification is a point-in-time assessment of controls, not a warranty of future performance or outcomes. A buyer should read the certificate as evidence of tested safeguards, not a guarantee of security.

AIUC-1 also publishes crosswalks to ISO 42001, the NIST AI RMF, MITRE ATLAS and the OWASP Top Ten, among others. If your customers already ask about those frameworks, the crosswalks show how AIUC-1 evidence relates to them.

Frequently asked questions