OWASP/AISVS

OWASP Artificial Intelligence Security Verification Standard (AISVS)

What it is

• A community‑driven catalogue of testable security requirements for AI‑enabled systems. It mirrors the OWASP Application Security Verification Standard (ASVS) but focuses on AI‑specific threats across the whole model lifecycle – from data collection and training to deployment, monitoring, and retirement.

Who it’s for

• Developers, architects, security engineers, auditors, and procurement teams that need a concrete checklist to verify that an AI application meets a defined level of security assurance.

How it’s organized

• Requirements are grouped into 12 chapters (e.g., Training Data Integrity, Input Validation, Model Lifecycle Management, Adversarial Robustness, etc.).
• Each requirement has a unique identifier like C9.4.3 and a verification level (1 = baseline, 2 = standard, 3 = high‑assurance).
• A Research Wiki accompanies every requirement, describing the threat it mitigates, verification approaches, tooling suggestions, and open gaps.

How to use it

1. Design phase – Treat the AISVS chapters as a security‑by‑design checklist when architecting AI pipelines.
2. Development / CI‑CD – Embed the relevant controls into code reviews, automated tests, and deployment pipelines.
3. Assessment – Use the standard as a framework for penetration testing, audits, or third‑party evaluations.
4. Procurement – Reference specific AISVS controls when evaluating vendors or off‑the‑shelf models.

Verification levels

• Level 1 – Baseline controls for any AI system (including internal tools).
• Level 2 – Standard controls for production systems handling sensitive data or making consequential decisions.
• Level 3 – Advanced controls for high‑assurance, safety‑critical, or regulated environments.

Current release

• The latest stable version is AISVS 1.01 (available as a PDF and as Markdown source). The repository follows a locked‑folder versioning scheme (1.0/, 1.01/, 1.02-dev/).

What it is not

• It is not a governance or risk‑management framework (those are covered by NIST AI RMF, ISO/IEC 42001, EU AI Act, etc.).
• It does not endorse specific tools or products.

Contributing

• Community members can open issues or submit pull requests to improve requirements, add references, or clarify verification steps. Security issues are handled via the project’s security policy.

License

• All content is released under the Creative Commons Attribution‑ShareAlike 4.0 International license.

---

In short, AISVS provides a practical, test‑oriented checklist for securing AI applications, helping teams move from vague best‑practice advice to concrete, auditable controls.