About and methodology

What it is, and what it is not

The tool is for teams starting from near zero. It shows you quickly what to do to get to a much better place. It covers the majority of what matters for common kinds of AI system without being comprehensive.

It is a starting point, not a complete solution. It does not:

• make you an expert in AI security;
• cover every permutation of AI system, or every risk to yours;
• adjust for your organisation: your sector, contracts, legal duties or the controls you already have;
• replace a security professional. If a failure would be serious, have someone qualified review your system.

How filtering works

1. You tick what your AI system does. If nothing is ticked, no requirements apply, so the assessment asks for at least one item.
2. Each item you tick adds the requirements mapped to it, taken from the OWASP AI Security Verification Standard (AISVS). Some items bring in others: an agent, a document search or a memory also generates content, so it gets those requirements too, and connecting to MCP servers makes it an agent.
3. Once you tick anything that describes an AI system you run, a small set of core requirements is added, such as pre-deployment testing, access control and logging. Providing an MCP server or using AI coding tools does not add them by itself.
4. Requirements are filtered based on the level you choose. Levels are cumulative: Level 2 includes all Level 1 requirements, and Level 3 includes everything.
5. Duplicates are removed and the list is sorted by requirement ID, then grouped by AISVS chapter.

You choose the level

The tool does not calculate a risk score. You pick the level of assurance your system needs, and each level says what kind of risk it usually suits. The levels are the three AISVS levels.

• Level 1, Essentials: the minimum any production AI system should have. Usually enough when the worst a failure could cause is minor inconvenience and no sensitive data is involved. A prototype or experiment may not need all of it.
• Level 2, Standard: the essentials, plus deeper defences, testing and monitoring. Usually right when a failure could expose personal, confidential or regulated data, or cause financial loss or legal harm.
• Level 3, Advanced: everything in the standard, for the highest assurance. Usually right when a failure could harm someone's safety, disrupt critical services, or expose highly sensitive personal data.

Level 1 is not a quick checklist: AISVS puts work such as testing before each deployment and watching for drift in inputs at Level 1. If you are unsure, start at Level 2 and adjust it later. You can change it on the results page at any time.

Limitations

• The mapping is a judgement about which requirements usually matter for each kind of system. It may include requirements you can rule out, and it may miss ones that matter in your context.
• The questions are broad. Two systems that tick the same boxes get the same list, however different they are underneath.
• Meeting every requirement on your list does not make a system bulletproof. It means the common gaps are closed.
• Only OWASP AISVS is used, no other framework or standard.
• AISVS covers what is specific to AI. General application security, such as sign-in, encryption, secrets handling and what you agree with a model provider about your data, is outside it and is not listed here.

Privacy

All logic runs in your browser. Your answers are kept only in this tab's session storage so they survive moving between pages. They are cleared when you start over or close the tab. If you copy a link to your list, your answers are written into the part of the link after the #, which browsers do not send to a server; anyone you give the link to can read them. Nothing is sent to a server, and there are no cookies, analytics or third-party scripts.

Disclaimer

This tool gives a security baseline to work from. It is not a compliance certification, audit, formal risk assessment, legal advice or a substitute for professional review, and it does not guarantee coverage of every risk to your system.

Who is behind it, and feedback

howtosecure.ai is put together by Keith Sinclair. The requirements are OWASP's: the site is essentially a wrapper around the AISVS that helps you find the requirements relevant to your system, and understand how to implement them.

If a mapping looks wrong, something is missing or unclear, or you have any other feedback, please message me on LinkedIn.

Attribution and licence

Requirement text is from the OWASP Artificial Intelligence Security Verification Standard (AISVS) v1.01 by the OWASP Foundation and AISVS contributors, licensed under Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0).

The data files on this site, including the scoping logic derived from AISVS, are shared under the same licence. howtosecure.ai is not affiliated with or endorsed by OWASP.