Free checklist
AIUC-1 readiness checklist
Every active AIUC-1 requirement in plain English, whether it applies to your agent, and the evidence a reviewer will ask to see. For founders and CTOs preparing for an AIUC-1 audit, or for an enterprise buyer’s AI security review that asks AIUC-1 style questions. Free to use, no sign-up.
As of October 16, 2026 · AIUC-1 as released on October 15, 2026 (the Q4 2026 version) · AIUC’s next release is due January 15, 2027
Not affiliated. Autogeny is an independent firm. We are not affiliated with, partnered with, endorsed by or accredited by the Artificial Intelligence Underwriting Company (AIUC). AIUC-1 is AIUC’s standard. Using this checklist does not count toward AIUC-1 certification. Check the official site, standard.aiuc-1.com, for the current version.
How to use it
It is a readiness check, not a certification, and not an auditor’s gap assessment. The requirement descriptions are our own short summaries. Follow the links for AIUC’s official wording.
- Pick one agent. Note what it can do: does it serve external users, take actions through tools, or generate code? Does it produce text, voice or images?
- Work through each table and mark it yes, partly or no. Print this page, or copy the tables into a spreadsheet. If a row does not apply to your agent, write down why. An auditor will ask.
- For each mandatory row marked no or partly, name an owner and a target date.
- Use the lists near the end as your evidence folder and test-readiness plan.
- Check the official site again before an audit starts. AIUC-1 changes every quarter.
This is general information, not legal advice.
AIUC-1 at a glance
- 50 active requirements in six areas: data and privacy (A), security (B), safety (C), reliability (D), accountability (E) and society (F) (AIUC: Scoping). 42 are mandatory and 8 are optional, by our count from the six domain pages.
- Some mandatory requirements depend on what your agent can do, such as generating code or taking actions. If your agent lacks that capability, the auditor can mark the requirement not applicable, with a short written reason. (AIUC: Requirements in scope)
- Each requirement has core controls; most also have optional extras. If you meet a requirement in a different way, you can offer alternative evidence for the auditor to judge. (AIUC: Scoping)
- Controls your customers can switch on themselves can count, if you document that in public docs, in the product or in onboarding. Secure defaults help. (AIUC: Fieldwork)
What changed on October 15, 2026
- New mandatory requirements: A009 (third-party testing for data and privacy risk) and E018 (third-party testing for accountability risk).
- B001, C010 and D004 now cover security, safety and reliability risk broadly, as defined in AIUC’s risk taxonomy.
- Retired: C011 and C012 (folded into C010) and D002 (folded into D004).
- D003 now expects human approval before sensitive tool actions, such as payments, deletions or permission changes.
- E006 now covers due diligence on the models themselves, not only the vendors.
- New optional controls for open-weight models (C002.4, E006.2, F001.3).
Source: AIUC-1 changelog.
How AIUC-1 certification works
| Question | Short answer | Source |
|---|---|---|
| Who tests the agent? | AIUC runs the technical evaluations itself today. Auditor-led evaluations are in development. A typical scope is 1,000 to 5,000 test scenarios per agent. TechCrunch reported “some 5,000 tests”. | Evals, Eval scope, TechCrunch |
| Who audits? | An AIUC-accredited auditor reviews your evidence and gives a verdict on each requirement. AIUC lists the accredited firms on its site. | Accredited auditors |
| Who issues the certificate? | Only AIUC. | FAQ |
| What does it take to pass? | No open critical (P0) or major (P1) findings in the evaluations, and a pass on every requirement that applies. | Evals |
| How long does it take? | AIUC’s own pages say 4 to 8 weeks, 5 to 10 weeks, and 5 to 12 weeks from kickoff. Expect longer if you start from zero. | Certification, FAQ, Auditor guide |
| How long is it valid? | 12 months. AIUC re-runs the evaluations every quarter. An accredited auditor re-audits each year, starting about two months before expiry. | FAQ, Re-certification |
| Which version applies? | The latest, with up to 30 days’ grace after a new release. The version is fixed at kickoff. | Requirements in scope |
| Point in time or over a period? | Point in time. There is no observation window like a SOC 2 Type II. | Fieldwork |
The 50 requirements
Six areas: 42 requirements are mandatory and 8 optional. Where the Mandatory? column starts with “If”, the requirement is mandatory when your agent has that capability; if it doesn’t, the auditor can mark it not applicable with a short written reason. Each ID links to AIUC’s official wording.
A. Data and privacy
9 requirements, 9 mandatory · Official page: standard.aiuc-1.com/data-and-privacy
| ID | Requirement, in plain words | Mandatory? | What a reviewer will ask to see |
|---|---|---|---|
| A001 | Tell customers how you use, keep and delete the data they put in, and enforce it. | Yes | Terms, privacy policy or DPA covering training use, inference use, retention and customer rights. The job or setting that deletes data on schedule. |
| A002 | Tell customers who owns AI outputs, how they may be used, and how to opt out or delete them. | Yes | Terms, DPA or AI addendum covering output ownership, use, opt-out and deletion. |
| A003 | Limit the data the agent can reach to what the user, role and task need. | Yes | Code or config that scopes retrieval and data access by user, role, session or task. |
| A004 | Stop the system leaking your company’s confidential information. | Yes | Guidance telling users what confidential information not to enter. Optional: provider no-training terms, leak detection. |
| A005 | Keep each customer’s data away from other customers. | Yes | Tenant isolation in databases and vector stores. Terms that disclose any combined use of customer data, with an opt-out. |
| A006 | Keep personal data out of outputs and logs. | Yes | PII detection, output filtering and log redaction in code or config. |
| A007 | Reduce the risk that outputs infringe copyrights or trademarks. | If external-facing | Model provider terms showing IP protections or indemnity. Optional: output filters, user notices. |
| A008 | Keep keys, tokens and passwords out of inputs, generated code, logs and storage. | If it generates code | Secret scanning on user input with a set response. Steps that keep secrets out of generated code. Credentials held in a secret manager. |
| A009 | Have an expert third party test data and privacy risk. New in Q4 2026. | Yes | Evaluation report at least every 3 months: risks tested, method, findings. See the third-party section below. |
B. Security
10 requirements, 7 mandatory, 3 optional · Official page: standard.aiuc-1.com/security
| ID | Requirement, in plain words | Mandatory? | What a reviewer will ask to see |
|---|---|---|---|
| B001 | Have an expert third party test security risk. Widened in Q4 2026. | Yes | Evaluation report at least every 3 months covering risks such as prompt injection and insecure outputs. |
| B002 | Detect and respond to prompt injection and other attacks in user input. | Optional | Detection rules and alerts, an incident log, and quarterly reviews of the rules. |
| B003 | Control what you publish about your AI system’s technical details. | Optional | Disclosure guidelines with an approval step. |
| B004 | Stop people probing or scraping your AI endpoints. | Yes | Rate limits and quotas, anomaly detection, an external (third-party) penetration test of the AI endpoints, and tickets showing fixes. |
| B005 | Screen user inputs in real time with a moderation tool. | Optional | Moderation settings: categories, thresholds, and block or warn actions. |
| B006 | Stop the agent acting beyond its intended scope and permissions. | If it takes actions | Allowlists or gateway rules on which services, APIs and MCP servers the agent can reach. Logs and alerts on agent actions. |
| B007 | Control who can see or change the AI system’s configuration, data, tools and prompt logs. | Yes | Role-based access settings. Access review records every quarter. |
| B008 | Secure the environment the AI system runs in. | Yes | Caller authentication on model APIs and agent interfaces, TLS on every connection, and credential rotation. |
| B009 | Limit how much the system returns at once, to reduce leakage. | If it outputs text, voice or images | Output limits (tokens, characters, results or time) in config or a demo. |
| B010 | Make generated code use secure patterns by default. | If it generates code | Prompts or config that steer to safe queries, output encoding, vetted auth and pinned, real packages. Sample outputs. |
C. Safety
10 requirements, 7 mandatory, 3 optional · Official page: standard.aiuc-1.com/safety
| ID | Requirement, in plain words | Mandatory? | What a reviewer will ask to see |
|---|---|---|---|
| C001 | Write down the AI risks that matter for your agent, with severity levels. | Yes | Risk taxonomy with categories, severity and examples for your use, tied to an outside framework. A record of the yearly review. |
| C002 | Test changes before release and get sign-off. | Yes | Test results with severity, a risk assessment of the issues found, and sign-off by the accountable person. |
| C003 | Prevent harmful outputs: distressing, hostile, offensive, biased or deceptive content, and risky advice. | If it outputs text, voice or images | Harmful-output filters or moderation settings. Guardrails on high-risk advice. |
| C004 | Keep the agent to its intended topics. | If it outputs text or voice | Guardrails or system prompt defining allowed topics. Logs of out-of-scope attempts. |
| C005 | Prevent the high-risk outputs specific to your agent. | Yes | Detection tied to your risk taxonomy, with an action (block, flag or log) for each severity. |
| C006 | Stop outputs carrying attacks to users, such as scripts or unsafe links. | If it outputs text, voice or code | Output sanitization (encoding, URL and schema checks). Handling or labels for untrusted content. |
| C007 | Flag high-risk outputs for a person to review. | Optional | Criteria for what counts as high risk, and the detection rules. |
| C008 | Monitor outputs across your risk categories. | Optional | A monitoring dashboard, output sampling logs or regular evaluation reports. |
| C009 | Let users stop, pause, redirect and give feedback in real time. | Optional | Working stop, pause and feedback controls that meet basic accessibility needs. |
| C010 | Have an expert third party test safety risk. Widened in Q4 2026. | Yes | Evaluation report at least every 3 months covering harmful, inappropriate and out-of-scope outputs and your own defined risks. |
C011 and C012 were retired on October 15, 2026. Their scope moved into C010.
D. Reliability
3 requirements, 3 mandatory · Official page: standard.aiuc-1.com/reliability
| ID | Requirement, in plain words | Mandatory? | What a reviewer will ask to see |
|---|---|---|---|
| D001 | Reduce made-up or wrong answers. | If it outputs text, voice or code | Checks of answers against source material. Citations or sources shown to users. |
| D003 | Stop tool calls doing unauthorized, out-of-scope or risky things. | If it takes actions | Tool allowlists and parameter checks, rate limits and caps, tool call logs, and human approval for sensitive actions. |
| D004 | Have an expert third party test reliability risk. Widened in Q4 2026. | Yes | Evaluation report at least every 3 months covering hallucinations, source attribution and wrong tool actions. |
D002 was retired on October 15, 2026. Its scope moved into D004.
E. Accountability
16 requirements, 14 mandatory, 2 optional · Official page: standard.aiuc-1.com/accountability
| ID | Requirement, in plain words | Mandatory? | What a reviewer will ask to see |
|---|---|---|---|
| E001 | Have a plan for AI privacy and security breaches. | Yes | A plan naming a breach lead, with notification steps, fixes (including freezing the system) and evidence preservation. |
| E002 | Have a plan for harmful outputs that seriously hurt a customer. | If external-facing | A plan with immediate steps (freeze, suppress outputs), named staff, and executive approval of customer communications about significant incidents. |
| E003 | Have a plan for wrong outputs that cause customers real financial loss. | If external-facing | A plan with the same parts as E002, focused on hallucinations. |
| E004 | Decide which AI changes need approval, name an approver for each, and keep records. | Yes | A change approval policy (model, system prompt, guardrails, user workflow, big swings in eval results) and sign-off records. |
| E005 | Document how you store data securely. | Yes | A storage security record covering sensitivity, regulation, controls and hosting choices. A trust center page can help. |
| E006 | Check your model providers and the models you use before you rely on them. | Yes | Assessment records: criteria (data handling, PII, security, compliance), scores, checks done, and the approval. |
| E008 | Review key AI decisions every quarter. | Yes | Quarterly review notes, a decision log, and a risk register that tracks fixes. |
| E009 | Log third-party access to your systems. | Yes | Logs of third-party API connections, sessions and data access, showing who, when and what. |
| E010 | Have an acceptable use policy for the AI, and enforce it. | Yes | The policy, detection of violations, and the message users see when they break it. |
| E011 | Record where AI data is processed. | Yes | Subprocessor list, cloud regions and inference endpoints, or a data flow diagram with a review date. |
| E012 | Track the AI and data laws that apply to you and how you comply. | Yes | A compliance register with your approach to each law, reviewed every 6 months. |
| E013 | Run a quality management system sized to your company. | Optional | Quality goals and metrics, change management, and issue tracking with root cause analysis. |
| E015 | Log what the AI system does. | Yes | Logging of inputs, outputs, steps and metadata. Log retention, access control and PII masking. |
| E016 | Tell users when they are dealing with AI. | Yes | Disclosure in each mode you use (chat, voice, AI actions, generated media) and a truthful answer to “are you AI?”. |
| E017 | Keep transparency documents for your AI system. | Optional | A model card, datasheet or AI bill of materials: versions, models, data sources, fine-tuning history and tool connectors. |
| E018 | Have an expert third party test accountability risk. New in Q4 2026. | Yes | Evaluation report at least every 3 months covering risks such as getting the agent to hide that it is AI. |
E007 and E014 were retired earlier and merged into E004 and E017.
F. Society
2 requirements, 2 mandatory · Official page: standard.aiuc-1.com/society
| ID | Requirement, in plain words | Mandatory? | What a reviewer will ask to see |
|---|---|---|---|
| F001 | Guard against the AI being used for cyberattacks. | If it outputs text, voice or code, or takes actions | Each model provider’s published cyber capability testing and mitigations, such as a model card. Optional: your own filters. |
| F002 | Guard against misuse for chemical, biological, radiological or nuclear harm. | If it outputs text, voice or images | Each model provider’s published testing and mitigations for these risks, for every model you use. |
The five quarterly third-party evaluations
Five mandatory requirements ask you to have expert third parties evaluate your agent at least every 3 months, against the risks in AIUC’s risk taxonomy. Society (F) has none.
| ID | Risk area | Examples of what gets tested |
|---|---|---|
| A009 | Data and privacy | IP infringement; leaks of personal data or credentials; leaks of business data such as the system prompt |
| B001 | Security | Unauthorized actions; unsafe links or code in outputs; insecure actions; prompt injection and jailbreak attempts |
| C010 | Safety | Physical or emotional harm; inappropriate or out-of-scope content; bias; medical, legal or financial advice; risks you define for your agent |
| D004 | Reliability | Hallucinations; misleading citations; wrong, unprompted or incomplete tool actions |
| E018 | Accountability | Getting the agent to hide or deny that it is AI |
What a reviewer will expect from each evaluation
- Who tested, with a record of their qualifications and their independence from your team.
- A scope and method based on the risk taxonomy, focused on your highest risks.
- Findings, and what you did about them (our advice; AIUC asks for the findings).
- A new round at least every 3 months.
Good to know
- In a certification audit today, AIUC runs these evaluations and supplies the results for the audit report. Testing earlier, by your own team or an outside firm, helps you find and fix problems first. It does not replace AIUC’s evaluations. (AIUC: Evidence collection, AIUC: Evals)
- After certification, AIUC re-tests the agent every quarter to keep the certificate valid. (AIUC: Maintaining certification)
- Other recurring work: B007 access reviews and E008 internal reviews are quarterly. AIUC also lists a 3-month cycle for C006 and for the optional B002 and C009. E012 is every 6 months. Most other requirements are yearly.
Get your agent ready to be tested
AIUC’s guidance on evaluation scope says access problems are the most common cause of delay. Have these in place before testing starts. (AIUC: Eval scope)
- A test instance that matches production: same configuration and guardrails, realistic seeded data, and tools that really run.
- API access to the agent, with long-lived or refreshable tokens, raised rate limits and enough model credits for thousands of test conversations.
- An architecture diagram, API docs and product docs that describe intended behavior.
- A list of every tool and data source the agent can use, marked read or write.
- Known correct answers for typical questions, so accuracy can be checked.
- Logs you can search that show tool calls and intermediate steps, not only final answers.
- A written list of what is out of scope (agents, features, tools, input types) and why.
AIUC notes that an agent reachable only through a user interface needs slower, more manual testing that costs more.
Documents and records to gather
Each item appears once, with the requirements it supports. “If” notes mark items that only apply to some agents. This list says what to have, not how to write it. AIUC says it has policy templates, and auditors can refer you to firms that help (AIUC: FAQ).
Policies and terms
- Customer data terms, in your terms of service, privacy policy, DPA or AI addendum: how inputs are used for inference and training, how long you keep them, and customer rights; who owns outputs and how they may be used; opting out and deleting outputs; disclosure and opt-out for any combined use of customers’ data. (A001, A002, A005)
- AI acceptable use policy, standalone or in your terms. (E010)
- User guidance on what confidential information not to enter. (A004)
- Change approval policy: which AI changes need sign-off, and who approves each. (E004)
- AI risk taxonomy with severity levels and examples. (C001, C005) If you use human review, your criteria for flagging high-risk outputs. (C007, optional)
- AI failure plans for privacy or security breaches (E001), harmful outputs (E002, if external-facing) and hallucinations that cause financial loss (E003, if external-facing). These can be sections of your incident response plan.
- Regulatory compliance register, reviewed every 6 months. (E012)
- Data storage security record. (E005)
- AI processing locations record. (E011)
Vendor and model records
- Due diligence file on each model provider and model: criteria, scores, checks and approval. (E006) For open-weight models, source, integrity and license checks. (E006, optional control)
- Model provider terms showing IP protections or indemnity. (A007, if external-facing)
- Each model provider’s published testing and mitigations for cyber misuse and for chemical, biological, radiological and nuclear misuse, such as model cards. (F001, F002)
Recurring records
- Pre-deployment test results, risk assessment and sign-off for each change that needs approval. (C002)
- Approval records for AI changes, such as pull request approvals or tickets. (E004)
- Yearly risk taxonomy review record. (C001)
- Quarterly internal review notes, decision log and risk register. (E008)
- Quarterly access reviews for AI configuration, data, tools and prompt logs. (B007)
- Logs of out-of-scope requests and any changes made in response. (C004)
- Third-party evaluation reports, at least every 3 months. (A009, B001, C010, D004, E018)
Security testing
- External (third-party) penetration test report covering your AI endpoints, and tickets showing the findings were fixed. (B004)
Optional, if you opt in
- Adversarial input incident log and quarterly detection rule reviews. (B002)
- Technical disclosure guidelines. (B003)
- Quality management documents. (E013)
- Model card, datasheet or AI bill of materials. (E017)
Technical evidence to have ready (config, code excerpts, screenshots or a demo)
- Data retention and deletion jobs. (A001)
- Data access scoping, tenant isolation, PII filtering and log redaction. (A003, A005, A006)
- Secret scanning, secret-safe code generation and credential storage. (A008, if it generates code)
- Rate limits and anomaly detection on AI endpoints. (B004)
- Agent service allowlists and monitoring. (B006, if it takes actions)
- Access controls on AI configuration, data, tools and prompt logs. (B007)
- Authentication and TLS on model APIs and agent interfaces. (B008)
- Output limits. (B009)
- Secure coding defaults. (B010, if it generates code)
- Harmful-output filters, advice guardrails, topic guardrails and risk-based detection. (C003, C004, C005)
- Output sanitization and labels for untrusted content. (C006)
- Groundedness checks and citations. (D001)
- Tool allowlists, parameter checks, rate caps, tool call logs and human approval steps. (D003, if it takes actions)
- Third-party access logs. (E009)
- Acceptable use detection and the message users see. (E010)
- Activity logging with retention, access control and PII masking. (E015)
- AI disclosure notices and the answer to “are you AI?”. (E016)
Sources
All AIUC pages accessed October 16, 2026.
- AIUC: AIUC-1 standard
- AIUC: AIUC-1 changelog (October 15, 2026 release)
- AIUC-1 domain pages: data and privacy, security, safety, reliability, accountability, society
- AIUC: Risks (risk taxonomy)
- AIUC: Third-party evals
- AIUC: Scoping the AIUC-1 audit
- AIUC: AIUC-1 Certification
- AIUC: Frequently asked questions
- AIUC: Re-certification: Maintaining AIUC-1
- AIUC: AIUC-1 Accreditation
- AIUC: AIUC-1 certification (auditor guide)
- AIUC: Agent under test and evaluation scope
- AIUC: Requirements in scope
- AIUC: Evidence collection
- AIUC: Technical evaluations (evals)
- AIUC: Auditor fieldwork
- AIUC: Maintaining certification
- Julie Bort. “Early Anthropic hire, former METR COO have found a way to rein in rogue AI agents.” TechCrunch, September 15, 2026
AIUC-1 is AIUC’s standard and its content belongs to AIUC. This checklist uses requirement IDs, our own short summaries and links. For the official requirements, evidence and wording, use AIUC’s site.
Want an independent readiness check?
Autogeny tests your agent as it runs, in a test copy of your own setup, maps the results to the requirements testing can evidence, and lists what else each remaining requirement needs. It is not an AIUC-1 audit or certification. Scope and prices are on the audit page.