A new standard has appeared that is specifically about AI agents, and it is raising a reasonable question for anyone already working towards ISO 42001: is this a competitor, a replacement, or something else? The standard is AIUC-1, and the short answer is that it is something else. It sits at a different level from ISO 42001 and is designed to work alongside it. If ISO 42001 governs how your organisation manages AI, AIUC-1 tests whether a specific agent actually behaves. You can, and for agent builders often should, have both.
Here is what AIUC-1 is, how it differs from ISO 42001, where each fits, and how to think about them together.
What AIUC-1 is
AIUC-1 is a security, safety and reliability standard built specifically for AI agents, the systems that take actions rather than just produce text. It is published by AIUC, the Artificial Intelligence Underwriting Company, which launched in 2025 and pairs the standard with independent audits and insurance. The shorthand people reach for is “SOC 2 for AI agents”, and it is a fair one.
What it certifies is a specific agent in a specific deployment, not an organisation. The requirements are organised into six pillars:
- Data and privacy (data leakage, personal data, IP)
- Security (prompt injection, jailbreaks, unauthorised actions)
- Safety (harmful or out-of-scope behaviour, human oversight)
- Reliability (hallucinations, unsafe tool execution)
- Accountability (ownership, change management, logging, vendor due diligence)
- Society (misuse and broader risk)
Underneath those pillars sit roughly 51 requirements and around 130 controls, scoped to what a given agent can actually do. The part that makes AIUC-1 distinct is how conformity is proven: not just documentation, but thousands of adversarial simulations, jailbreaks, injection attacks, data-leakage attempts, hallucination probes, run against the agent, followed by an independent third-party audit. Certification lasts a year, but the agent is re-tested quarterly, and the standard itself updates on a quarterly cadence. Recent 2026 releases, for example, added controls for Model Context Protocol (MCP) and agent-to-agent communication, agent identity, and third-party risk, which tells you how quickly this space is moving.
Because AIUC also underwrites, certification feeds insurance: safer agents, better terms.
What ISO 42001 is, briefly
ISO/IEC 42001 is the international standard for an AI management system (AIMS). It certifies the organisation: whether you have built and operate a system of governance, roles, risk assessment, lifecycle controls, oversight, evidence and continual improvement across a defined scope. It is the AI equivalent of ISO 27001 for information security, and it is the cleanest way to demonstrate to a customer, auditor or regulator that you govern AI responsibly as an organisation. We cover what a working AIMS contains separately.
The important thing for this comparison: ISO 42001 proves you have the system. It does not, by itself, prove that any particular AI agent you run actually works safely in production.
The core difference: two different levels
This is the whole point, so it is worth stating plainly.
| ISO 42001 | AIUC-1 | |
|---|---|---|
| What it certifies | The organisation’s AI management system | A specific AI agent in a specific deployment |
| The question it answers | Do you govern AI responsibly? | Does this agent behave safely under pressure? |
| Level | Organisation-wide | Product / use-case |
| Primary evidence | Policies, roles, risk process, records | Adversarial test results + audited controls |
| Cadence | Certification with surveillance audits | Quarterly re-testing, annual re-audit |
| Best at | Showing the system is in place | Showing the agent actually holds up |
ISO 42001 is the management system. AIUC-1 is the behaviour of the thing the management system is supposed to be governing. A clean ISO 42001 certificate tells a buyer you run AI well; it does not tell them your customer-service agent will not leak another customer’s data or invent a refund it is not allowed to give. AIUC-1 is built to answer exactly those questions. And the reverse holds too: an AIUC-1 pass is a snapshot of one agent at one time, which only stays meaningful if there is a management system keeping it current. Each covers the other’s blind spot.
Where they overlap (by design)
AIUC-1 was not built in isolation from ISO 42001. It incorporates most of the ISO 42001 control set and turns the management-system approach into concrete, testable requirements, then layers agent-specific testing, failure planning and system security on top. In particular, most controls in AIUC-1’s accountability domain map directly to ISO 42001.
The practical consequence is encouraging rather than duplicative: an organisation already certified to ISO 42001 arrives at an AIUC-1 assessment carrying a real head start on the governance requirements, and can focus its effort on the agent-level testing that is genuinely new. The two standards reinforce each other, they do not make you do the same work twice. The overlaps show up in the usual places, impact assessment, data governance, logging and incident response, foundation-model supplier due diligence, and user transparency, which are the same disciplines any serious AI assurance programme already cares about.
And where does SOC 2 fit?
A common point of confusion. SOC 2 attests to a service organisation’s general security controls. It tells a buyer the company running the system has sound security practices. It says nothing about whether an AI agent resists prompt injection or hallucinates a policy, and nothing about whether the organisation governs AI as a discipline. So the three are not competitors but a stack: SOC 2 for the company’s security baseline, ISO 42001 for the AI management system, and AIUC-1 for the behaviour of the agent itself. This is the same pattern we see with independent, third-party assurance generally, different instruments answering different questions.
Which do you actually need?
It depends on what you are and what you are being asked to prove.
- You deploy AI across your organisation and need to show it is governed. ISO 42001 is the more relevant standard. It gives you, and anyone assessing you, confidence that AI is managed as a discipline rather than tool by tool.
- You build or ship an AI agent and customers want assurance it is safe. AIUC-1 answers the agent-specific questions that ISO 42001 and SOC 2 cannot, with evidence (adversarial testing) that a questionnaire never could.
- You build agents and operate at any scale where trust matters. You will likely want both: the management system at the organisation level and tested assurance at the agent level. Together they are a far more complete story than either alone.
A few honest caveats are worth keeping in view. AIUC-1 is new (2025) and moving fast, which is a strength for relevance but means the target shifts quarterly. It is US-founded and commercially driven: the same company publishes the standard, runs audits and sells the insurance, which is worth understanding even if the development partners are credible. And it is agent-specific, not a general AI governance framework, so it complements but does not replace ISO 42001 or your obligations under the EU AI Act.
How this fits a governed build
The reason this comparison matters to us is that the two levels AIUC-1 and ISO 42001 sit at are exactly the two levels we already build at. A governed agent is one whose actions are bounded, observable, evaluated against a test suite, reversible and accountable, which is precisely the kind of evidence AIUC-1 puts to the test at the agent level. And the management system that keeps that true over time, the register, the risk process, the oversight, the records, is what ISO 42001 certifies at the organisation level.
So whether or not you pursue an AIUC-1 certificate specifically, building agents that could stand up to it, inside an ISO 42001 management system, is simply what responsible agent delivery looks like now. The standards are catching up to a discipline, not inventing one.
The short version
AIUC-1 and ISO 42001 are not competitors. ISO 42001 certifies your organisation’s AI management system, how you govern AI across the business. AIUC-1 certifies a specific AI agent, proving through adversarial testing that it actually behaves safely in production, and it is re-tested quarterly rather than assessed once. They overlap by design, AIUC-1 builds on most of the ISO 42001 control set, so the two reinforce each other: the management system at the organisation level, tested assurance at the agent level. SOC 2 sits underneath both as the general security baseline. If you deploy AI, ISO 42001 is your anchor. If you build agents, AIUC-1 answers questions nothing else does. If you do both, you want both.
When you want to build AI agents that are governed and defensible, inside a management system that holds up, that is the work we do: see our AI agent development and ISO 42001 consulting, or get a quick read on where your governance stands with the free AI governance readiness assessment. For the wider picture, start with what AI governance is.
Frequently asked questions
What is AIUC-1?
AIUC-1 is a security, safety and reliability standard built specifically for AI agents, the systems that take actions rather than just answer questions. Published by AIUC (the Artificial Intelligence Underwriting Company) and often described as "SOC 2 for AI agents", it certifies a specific agent in a specific deployment against a catalogue of requirements across six pillars: data and privacy, security, safety, reliability, accountability and society. Certification involves thousands of adversarial tests (jailbreaks, prompt injection, data leakage, hallucination) plus an independent third-party audit, and it is re-tested quarterly rather than assessed once.
How is AIUC-1 different from ISO 42001?
They operate at different levels. ISO 42001 certifies your organisation's AI management system: the governance, roles, risk process, lifecycle discipline and evidence that show you manage AI responsibly across a defined scope. AIUC-1 certifies a specific AI agent: it proves, through technical testing, that this agent resists prompt injection, does not leak data across tenants, does not invent policies under pressure, and behaves safely in production. ISO 42001 shows you govern AI well but not that any given agent works; AIUC-1 shows a given agent is safe but is a point-in-time snapshot that only stays meaningful with ongoing governance.
Do AIUC-1 and ISO 42001 overlap, and can you hold both?
Yes to both. AIUC-1 deliberately incorporates most of the ISO 42001 control set and turns the management-system approach into concrete, testable requirements, then adds third-party testing, AI-specific failure plans and agent security on top. Most controls in AIUC-1's accountability domain map directly to ISO 42001, so an organisation already certified to ISO 42001 walks into an AIUC-1 assessment with a real head start on the governance side. Holding both is the most complete position: the management system at the organisation level, and tested assurance at the agent level.
Who needs AIUC-1?
Primarily organisations that build or ship AI agents and need to give customers credible, agent-specific assurance, and the enterprises buying those agents who want more than a vendor questionnaire. If you deploy AI broadly across your organisation and need to demonstrate governance, ISO 42001 is the more relevant standard. If you build agents that act in production, AIUC-1 answers the questions ISO 42001 and SOC 2 cannot: does this specific agent actually hold up under adversarial pressure? Many agent builders will want both.
Does AIUC-1 replace ISO 42001 or SOC 2?
No. SOC 2 attests to a company's general security controls; it says nothing about whether an AI agent resists prompt injection or hallucinates a refund policy. ISO 42001 proves the organisation governs AI well but not that a particular agent is safe. AIUC-1 fills that specific gap at the agent level. They answer different questions and are strongest together: SOC 2 for the company's security baseline, ISO 42001 for the AI management system, AIUC-1 for the behaviour of the agent itself.
How often is an AIUC-1 certified agent re-tested?
Quarterly. AIUC-1 is explicitly not a one-time snapshot: certified agents are re-tested against thousands of adversarial scenarios every quarter, and the full technical and legal controls are re-audited annually. The standard itself also updates on a quarterly cadence, with recent 2026 releases adding controls for Model Context Protocol (MCP) and agent-to-agent communication, agent identity and access, and third-party risk, reflecting how fast agent architectures are moving.
Building something you need to govern?
Start with a fixed-scope AI Opportunity & Risk Audit.
Meet an Expert