Passing the Security Review Your Buyers Run in the AI Era
Enterprise buyers now ask how your SaaS product is built, including how you govern AI-written code. Here is the evidence they check and how to have it ready.

To pass a vendor security assessment today, a B2B SaaS company has to prove three things with evidence: that its controls operate as its policies say, that its product is secure by design, and that it governs how AI is used to build and run the product.
A vendor security assessment is the review an enterprise customer runs before it lets your product touch its data. For years it was mostly paperwork: a questionnaire, a certificate, a penetration test report. That is changing. Since the US Cybersecurity and Infrastructure Security Agency (CISA) and the FBI published the Secure by Demand Guide in August 2024, buyers have been pushed to ask how a product is engineered, not only how the company is run. And as AI coding tools write more of the code in the products under review, buyers now ask a newer question: who checks what the AI wrote before it ships?
This guide is for the CTO, VP of Engineering, head of product or security lead at a B2B SaaS company that sells to enterprises and has a deal sitting in security review.
What a vendor security assessment checks now
A vendor security assessment checks whether your company and your product can be trusted with the buyer’s data and users. In practice it has three layers: a questionnaire about your security program, a request for evidence that backs the answers, and a growing set of questions about the product itself, including how AI is used to build and operate it.
The questionnaire is the visible part. Many enterprise buyers send a standard template rather than writing their own. The Shared Assessments Standardized Information Gathering (SIG) questionnaire measures risk across 21 domains, including access control and Artificial Intelligence. The Cloud Security Alliance’s CAIQ is the other common template, especially for cloud services. Some buyers send their own spreadsheet built from internal policy.
Whatever the format, the domains repeat. When we work through a buyer’s questionnaire with a client, these are the areas it asks about:
- Security policies and controls. Whether a documented information security program exists, and who owns it.
- Data protection. How customer data is encrypted, segregated, retained and deleted.
- Access control. Who can reach production and customer data, how access is granted and reviewed, and whether multi-factor authentication is enforced.
- Incident response. How you detect, contain and notify, and how fast.
- Privacy. Lawful processing, data processing agreements and sub-processors.
- Third-party and supplier oversight. Which services you depend on, and how you assess them.
- Business continuity. Backups, recovery objectives and whether they have been tested.
The second layer is evidence. A buyer’s security team does not take “yes” on trust. It asks for the artifact behind the answer: the access review log, the encryption configuration, the incident runbook, the latest penetration test report and its remediation status, a certificate with its scope statement.
The third layer is the product. This is where reviews have changed most in the last two years, and it is the part SaaS teams prepare for least.
Why buyers moved from paperwork to product security
Buyers moved toward product security because government guidance told them to use their purchasing power to demand software that is secure by design. CISA’s Secure by Demand Guide gave every software buyer a ready-made list of product questions to ask before, during and after procurement.
CISA framed the problem plainly when it released the guide: acquisition teams often understand the core cybersecurity requirements for a purchase, but frequently don’t assess whether a supplier treats security as a core consideration from the earliest stages of product development. The guide is short, and it is written for any software customer, not only government agencies. That is why its questions now turn up in private-sector questionnaires and contract negotiations.
What the Secure by Demand Guide asks a software manufacturer is specific:
| Area | What the buyer is told to ask or collect |
|---|---|
| Authentication | Standards-based single sign-on at no additional cost; MFA or phishing-resistant options such as passkeys enabled by default, at no cost; default passwords eliminated |
| Classes of vulnerability | Which classes of defect have been systematically eliminated, such as SQL injection through parameterized queries, and a roadmap for the rest |
| Evidence of intrusions | Security logs in the baseline product; for SaaS, logs retained and available to customers for at least six months at no extra charge |
| Software supply chain | A software bill of materials (SBOM) in a standard, machine-readable format, covering all third-party and open source dependencies |
| Vulnerability disclosure | A published vulnerability disclosure policy and accurate CVE records |
For a SaaS company, several of these are product decisions, not policy documents. Charging extra for SSO, shipping MFA as an opt-in, or keeping audit logs for 30 days in the default plan are all choices a buyer can now point at. If your answer to “is SSO included at no extra cost?” is “only on the Enterprise tier”, expect a follow-up.
The supply chain question has also grown. In May 2026, CISA and its G7 partners released Software Bill of Materials for AI: Minimum Elements, guidance that extends the SBOM idea to AI systems and is meant to be read in addition to the general SBOM minimum elements. If your product ships AI features, buyers now have an official reference for asking what models, datasets and providers sit inside it.
Regulation points the same way in Europe. Under the EU Cyber Resilience Act, reporting obligations for manufacturers have applied since 11 September 2026, and the main obligations apply from 11 December 2027. Whether the CRA applies to your product is a legal question for your counsel, but its direction is clear, and enterprise buyers in Europe are already writing its expectations into their reviews.
The practical consequence for a B2B SaaS company: the review is no longer only about whether your company runs a security program. It is about whether your product was built so that its customers don’t have to compensate for it. Buyers also know attackers now use AI to find weak products faster, as our analysis of whether AI can hack your software shows, which raises the bar for what they accept.
The new question: how do you govern AI-generated code?
Buyers now ask whether AI writes your code because independent testing still finds security flaws in a large share of AI-generated code, and because, in Linux Foundation research, 3 in 5 teams now ship code faster than their security reviews can keep up. The question in the questionnaire is usually simple (“Do you use AI coding tools?”), but the follow-ups are about control: who reviews AI-written changes, what checks run before merge, and what evidence shows it.
The data behind the question is recent. In Veracode’s Fall 2026 GenAI Code Security update, published on October 29, 2026, 18 models were tested across four languages. The average security pass rate rose from 56% to 62%, and the vulnerability rate fell from roughly 44% to about 38%. That is real progress. It still means the best models, which tied at a 73% pass rate, produced insecure code in more than a quarter of security-relevant tasks.
The review gap is the other half. The 3-in-5 figure comes from the Linux Foundation’s State of AI Impact on Software Security and Cybersecurity 2026, published in partnership with OpenSSF in November 2026. In the release announcing the study, ActiveState, one of its sponsors, reported that among teams that have an AI security policy, only 57% require human review of AI-generated code.
Put those two findings in front of an enterprise security team and the question writes itself. If the best models still fail more than a quarter of security tasks, and only 57% of teams with an AI security policy require a person to check the output, the buyer wants to know which kind of supplier you are.
The standard questionnaires have caught up. The SIG includes an Artificial Intelligence domain, and its published mappings include ISO/IEC 42001:2023 and NIST AI 100-1, the AI Risk Management Framework. The Cloud Security Alliance released its AI Controls Matrix v1.1 in June 2026, with 247 control objectives across 18 security domains, and its companion AI-CAIQ v1.1 contains 320 assessment questions. A buyer who wants a structured AI section no longer has to invent one.
What a good answer looks like is not “we don’t use AI”. For most SaaS teams that is either untrue or a competitive disadvantage. A good answer describes the controls:
- Name the tools and their access. Which AI coding assistants and agents your engineers use, and what each one can reach: repositories, tickets, test environments. State plainly whether any of them can access production or customer data.
- Keep a human decision on what ships. Every AI-written change goes through pull request review by an engineer who is accountable for it, with mandatory senior review on sensitive paths such as authentication, authorization, payments and data export.
- Run the same automated checks on every change. Static analysis, dependency and secret scanning, and tests in CI, applied whether a person or a model wrote the code.
- Govern what the AI pulls in. AI assistants add dependencies. In the same study, as OpenSSF summarized it, 60% of organizations and projects said AI tools had increased the number of external dependencies in their code. Your dependency policy and SBOM should cover them.
- Keep the record. Pull request history, CI logs and review approvals are the evidence. A policy without records is a claim; a policy with records is an answer.
This is the governance layer the review is checking. The engineering detail of how to secure AI-written code before it merges is covered in our guide to secure application development; the review only needs to see that the controls exist and operate.
Our own delivery work follows the same pattern. On a consumer marketplace platform we develop with Claude Code, the engineer reviews what Claude produces and decides what ships, tests are written alongside each change and run in CI, the client verifies every feature on staging before release, and Claude never touches real user data. On the platform’s messaging layer, the engineer ran a security review with Claude Code on every feature before it shipped. Each of those sentences maps to an artifact a buyer can inspect.
What evidence actually passes a review
Evidence passes a review when it shows a control operating over time, not just existing on paper. An access control policy is a starting point; the quarterly access review record that matches it is what a buyer’s security team accepts. The fastest way to fail is to answer “yes” to a question you cannot back with a dated artifact.
Most reviews stall in the same place: the answer and the evidence disagree. The questionnaire says MFA is enforced for all administrators, and the identity provider export shows two service accounts without it. The policy says logs are kept for a year, and the logging configuration keeps them for 30 days. Neither gap is a disaster on its own. Both tell the buyer that nobody has checked the answers against reality, and that is what triggers the second and third rounds of questions.
Map each common question to the artifact that settles it:
| Question area | Artifact that answers it |
|---|---|
| Access control | Identity provider export showing MFA enforcement; a dated access review record; a joiner-mover-leaver procedure with recent examples |
| Data protection | Encryption configuration for data at rest and in transit; key management description; data retention and deletion settings |
| Logging and monitoring | Logging configuration with retention period; an example of an alert and its handling |
| Vulnerability management | Latest penetration test report, its scope, and the remediation status of each finding; scanning results from CI |
| Secure development | Pull request policy, branch protection settings, CI security checks, and the AI coding policy described above |
| Incident response | Incident response plan, notification commitments, and evidence of a recent exercise or real incident review |
| Supply chain | SBOM, sub-processor list, and how new dependencies and providers are approved |
| AI features | Model and provider inventory, data flows into and out of models, prompt injection and data leakage testing results |
Two artifacts deserve special attention for SaaS teams.
The first is the penetration test report. Buyers read the scope as closely as the findings. A test that covered the marketing site and the login page says little about the multi-tenant API that holds their data. An independent penetration test scoped to the product surfaces the buyer cares about, with a dated remediation record, answers far more questions than a broader test of the wrong surface.
The second is the AI feature record. If your product ships an assistant, a summarizer or an agent, buyers will ask about prompt injection, data leakage through model responses and who can access the model under which permissions. Our breakdown of generative AI security risks covers where those vulnerabilities live in a product. For the review, what matters is a findings log: what was tested, what was found, and what was fixed.
The buyer is reading for consistency as much as for strength. A modest program with answers that match its evidence passes more often than an impressive one with gaps between the two.
Where ISO 27001 carries you, and where it stops
ISO 27001 certification carries a SaaS company a long way through a vendor security assessment, because it proves an audited information security management system with defined controls, owners and evidence. It stops at the edge of its scope and its design: it does not certify your product, it does not by itself establish compliance with regulations such as NIS2, and it was not written for AI-specific risks like prompt injection or model behavior.
What ISO 27001 gives you in a review is a shared language. A certificate tells the buyer that an independent auditor has examined your information security management system and the evidence that its controls operate. In many reviews, the certificate and its scope statement stand in for answers to whole sections of the questionnaire. ISO 27001 is accepted internationally, including in US security reviews, so one certificate can serve buyers in Europe and the US.
We know this from the inside. unicrew is ISO 27001:2022 and ISO 9001:2015 certified, audited by Quay Audit UK, and we renew through a multi-stage audit, so we demonstrate our own controls to an auditor on a fixed cycle. That experience shapes how we review a client’s evidence: we look for the artifact an auditor would ask for, not the paragraph a policy template would supply.
The limits matter just as much, and buyers know them:
- Scope. A certificate covers the scope in its statement. If your scope is “the corporate IT environment” and the buyer is assessing your production platform, the certificate answers less than you think. Check the scope statement before you rely on it.
- Organization, not product. ISO 27001 certifies a management system. It does not say your API is free of broken access control. That is why buyers still ask for penetration test reports and product security answers on top of the certificate. The same holds for our certification: it covers how unicrew operates, and it does not certify a client’s product.
- Regulation. ISO 27001 alone does not make an entity NIS2 compliant. Registration, incident reporting and management duties have to be assessed for the specific entity and jurisdiction. The same logic applies to GDPR processing requirements.
- AI. ISO 27001 controls cover secure development and supplier management in general terms. They don’t ask how you test a model for prompt injection, how you inventory the models and datasets in your product, or how you assess the impact of an AI system on the people it affects.
None of this makes ISO 27001 less valuable. It makes it the foundation rather than the finish. The honest position for a SaaS company shipping AI features in 2026 is that its information security management system has to be extended to cover AI-specific risk, and buyers increasingly ask how that extension was done.
Extending compliance to AI risk
You extend an ISO 27001-based program to AI risk by adding an AI governance layer: an inventory of AI systems and models, AI-specific risk and impact assessment, controls for AI development and operation, and evidence that those controls run. ISO/IEC 42001 provides the management system structure for this, the NIST AI Risk Management Framework provides a risk vocabulary many buyers use, and the CSA AI Controls Matrix turns both into assessable controls.
The three frameworks do different jobs, and a SaaS company does not need all of them at once.
ISO/IEC 42001:2023 gives AI governance a management system structure. Published in December 2023, it specifies requirements for establishing, implementing, maintaining and continually improving an AI management system, and ISO describes it as a Plan-Do-Check-Act management system standard rather than a set of rules for specific AI applications. For a company that already runs ISO 27001, that is the practical point: the AI management system can extend the policies, risk process, internal audit and management review you already operate, instead of standing up a parallel program.
The NIST AI Risk Management Framework, released on January 26, 2023 and intended for voluntary use, gives a review its shared vocabulary. Many buyers frame their AI risk questions in its terms, and the SIG maps to it, so answers that follow its structure are easier for a reviewer to accept. For a team that builds or fine-tunes models, NIST’s SP 800-218A (July 2024) adds AI-specific practices to the Secure Software Development Framework (SSDF).
The CSA AI Controls Matrix is where frameworks become questions. With 247 control objectives and a 320-question AI-CAIQ, it is the closest thing to a standard AI questionnaire a buyer can send. CSA maps the matrix to ISO 42001 and ISO 27001, so work done against one framework can be evidenced against the others.
How far to go depends on your product:
- AI used only to build the product. If AI is in your development process but not in the product, the main requirement is the AI coding governance described earlier: tool inventory, access limits, human review and records. This usually sits inside your existing ISO 27001 secure development controls, extended with an AI usage policy.
- AI features that touch customer data. If your product sends customer data to a model, you need the full product-side picture: a model and provider inventory, data flow documentation, prompt injection and data leakage testing, access control on model capabilities, and an incident process that treats model misbehavior as a security event. An AI SBOM is the natural place to record the inventory.
- AI as the core product. If the product is an AI system, buyers will expect an AI management system aligned with ISO 42001, and some will ask whether you are certified or working toward it.
Our own position here comes from running AI products as well as building them for clients. We build and operate Snaplore and Talkmetry, our own AI products, and the questions a buyer asks about model access and data handling are the questions we answer for ourselves. That is why AI application security is part of every readiness review we run where the product ships AI features.
How long readiness takes, and what to fix first
Readiness takes as long as the gap between what your answers claim and what your evidence shows. A team with ISO 27001 in place and a recent product penetration test may close a buyer’s questionnaire in days. A team starting without documented controls needs weeks to months, because some evidence, such as a record of quarterly access reviews, can only be produced by running the control over time.
The order of work matters more than the speed. When the questionnaire is already in your inbox, sequence the effort by what blocks the deal:
- Read the whole questionnaire first. Map every question to an owner and an artifact before answering anything. This shows the real gaps in an afternoon and stops you from writing answers you will later have to retract.
- Fix the visible technical controls. MFA enforcement on every administrative and service account, removal of shared credentials, logging retention that matches your policy. These are fast and they are the controls reviewers check first.
- Close the evidence gaps on what you already do. Many teams do the right thing and never recorded it. Export the access review, document the incident you handled last quarter, write down the AI coding policy your engineers already follow.
- Commission what you cannot produce yourself. A current penetration test scoped to the product, or an AI application security review if you ship AI features. Book these early: they take calendar time, and their findings need time to remediate.
- Plan what takes months. Certification, a mature vulnerability disclosure program, or an ISO 42001 extension. Say so honestly in the questionnaire, with a dated plan. Buyers accept a credible roadmap far more readily than a vague “yes”.
Three decisions keep this from repeating with every deal. Keep one evidence library, updated on a schedule, so the next questionnaire starts from current artifacts. Keep a single owner for security answers, so different deals don’t get contradictory responses. And treat the buyer’s findings as a product backlog: the same gaps that slow one enterprise deal will slow the next.
Who does the work is a real choice for a SaaS company. An in-house security engineer is right when security is a continuing responsibility you need inside the company. A compliance automation platform helps organize evidence once controls exist. An outside review is useful when you need an independent, evidence-led view against a specific buyer’s requirements on a deadline. And when the gaps are in the product itself, such as missing SSO, short log retention, an untested AI feature or a release pipeline without security checks, someone has to build the fix, which is an engineering capacity question as much as a security one.
When a readiness review is the wrong first step
A readiness review is the wrong first step when the product needs engineering work before its controls can be meaningfully assessed, or when the review you face calls for something a readiness assessment does not provide, such as active testing or a certification audit.
Some situations call for a different starting point:
- You already have an agreed remediation plan. If a previous review or audit produced a clear list of fixes, the next step is implementation, not another assessment. Our cybersecurity consulting team can implement the fixes you select in a separately scoped engagement.
- The product is unstable. If releases regularly break production or the codebase has no reliable tests, controls reviewed today will be out of date next sprint. Stabilize first, with software project rescue if the problems run deep.
- The buyer asked for a penetration test. A readiness review checks controls and evidence; it does not attack your product. If the buyer names a test, commission a penetration test scoped to the surfaces they care about.
- You need certification itself. A readiness gap review prepares you for a certification audit, but the certificate is issued by a certification body after its own audit. Readiness findings are not an attestation, and the buyer decides whether to approve.
- You are still deciding what AI belongs in the product. If the AI feature is not built yet, start with an AI readiness assessment to decide which use case is worth building; securing it comes after that decision.
Picking the right first step matters because a polished report that tells you what you already knew does not move the deal. When the gaps are already known, what moves it is engineers who fix them.
How we prepare a product for a buyer’s review
We prepare a SaaS product for a buyer’s security review the way an auditor would examine it: starting from the buyer’s actual questions, checking every answer against the artifact that proves it, and ranking what needs work by its relevance to the deal. Then, where the gaps are in the product, our engineers can build the fix inside the client’s own codebase and release process.
We work inside software vendors’ product organizations: companies that build and sell their own platforms, whose enterprise customers run exactly the reviews described here. Our engineers work within the client’s codebase, reviews and release process, so the security bar their buyers set becomes ours too.
Our approach rests on four habits drawn from our own audits and client delivery:
- Requirements before code. On a private online consultation platform in healthcare, our team studied the HIPAA Security Rule in depth before development started, and that study shaped access and audit controls, integrity and authentication. Security requirements that are settled at design time produce evidence naturally; requirements added after launch produce exceptions.
- Security built into the product, not around it. On a multi-factor authentication platform that uses 3D facial verification and machine learning to detect fraud, we refactored the platform’s data architecture, where a fraud-scoring decision sits in the middle of a login. Work like that sits in the product itself, which is where a buyer’s review of your authentication, MFA and SSO looks.
- Long-term platforms that keep pace with their rules. We have supported a document management SaaS platform since 2014, through regulations that have changed more than once. Keeping a growing platform compliant is continuous engineering work, not an annual project.
- AI-assisted delivery with a human decision on every release. As described above, our engineers review what AI produces and decide what ships, with tests, security review and staging verification on every feature. When a client’s buyer asks how AI-written code is governed, the client can point to records rather than intentions.
When a deal is already in review, our Security and Compliance Readiness Sprint is the fixed-scope route. We sign an NDA before any access, work from read-only, least-privilege access agreed with your technical contact, and review your code, infrastructure configuration and policies against the buyer’s questionnaire, ISO 27001 practices where in scope, and AI application security where the product ships AI features. You receive a control-by-control gap review, a report written for the buyer, auditor or board, a remediation roadmap ranked with deal blockers first, and an AI application security findings log. In total, the review takes between four and ten hours of your team’s time.
What happens after the report is your choice. Your engineers can implement the roadmap, or our cybersecurity consulting team can implement the fixes you select in a separately scoped engagement, from SSO and log retention to CI security checks. Where the work is mostly pipeline and infrastructure, our DevOps engineers handle logging, access control and environment separation. And for teams that need a senior engineer embedded in their delivery, a Forward Deployed Engineer works within your team, writes production code and takes responsibility for the outcome, with the same AI review discipline described above. Most engagements start within two to four weeks.
If a buyer’s questionnaire is already in your inbox, the fastest path is an evidence-led gap review against the buyer’s own questions. Tell us which product and review you are preparing for, the standards named and your deadline, and we will scope the work around them.
Key takeaways
- The review now covers the product, not only the company. CISA’s Secure by Demand Guide gave buyers specific product questions on SSO, MFA by default, log retention, SBOMs and vulnerability disclosure, and those questions now appear in enterprise reviews.
- “Do you use AI to write code?” needs a governed answer. Name the tools and their access, keep a human decision on what ships, run the same checks on every change, govern new dependencies, and keep the records.
- Evidence beats assertion. A modest program whose answers match dated artifacts passes more reliably than an impressive one with gaps between policy and practice.
- ISO 27001 is the foundation, not the finish. It proves an audited management system within its scope, but it does not certify your product, settle NIS2 on its own, or address AI-specific risk.
- Extend compliance to AI deliberately. Use ISO/IEC 42001 for the management system, the NIST AI RMF for risk vocabulary, and the CSA AI Controls Matrix and AI-CAIQ for assessable controls, sized to how much AI your product contains.
- Sequence readiness by what blocks the deal. Map the questionnaire, fix visible controls, document what you already do, commission tests early, and give buyers a dated plan for the rest.
Frequently asked questions
A vendor security assessment is the review an enterprise customer runs on a software supplier before signing, and often again at renewal. For a B2B SaaS company it usually combines a security questionnaire, a request for evidence such as policies, an ISO 27001 certificate or a recent penetration test report, and follow-up questions on how the product itself is built and operated. The buyer's security team decides whether the risk is acceptable and what goes into the contract.
A vendor security questionnaire is the structured set of questions a buyer sends to assess your security posture. Many buyers use a standard template, such as the Shared Assessments SIG or the Cloud Security Alliance CAIQ, and others write their own. The questions cover access control, data protection, incident response, vulnerability management, third parties and, increasingly, AI. Each answer should point to an artifact that proves it, not just a policy statement.
An ISO 27001 readiness assessment is a gap review of your controls and evidence against the standard before you face a certification audit or a buyer's review. It shows which controls are documented, which are actually operating, and which need work, ranked by risk. Our Security and Compliance Readiness Sprint includes this gap review where ISO 27001 is in scope, alongside the buyer's own questionnaire and, where the product ships AI features, an AI application security review.
Yes, increasingly. Standard questionnaires now carry AI sections: the Shared Assessments SIG includes an Artificial Intelligence domain, and the Cloud Security Alliance publishes an AI-specific questionnaire, the AI-CAIQ. Buyers want to know which AI tools touch your code and data, who reviews AI-generated changes before release, and what evidence shows that review happens. A written policy plus pull request and pipeline records answers it best.
Our readiness sprint has a fixed fee tied to an agreed scope, and both are confirmed in writing before work starts. The fee depends on the product and environments in scope and the requirements being reviewed, such as one buyer's questionnaire or a full ISO 27001 gap review. We quote it on a short scoping call once we have the questionnaire or the standards that apply.
It depends on the gap between what your questionnaire answers claim and what your evidence shows. If your controls already operate, most of the work is collecting the artifacts that prove them. Some evidence, such as a record of quarterly access reviews, can only be produced by running the control over time, so a control you have not yet started sets the floor. A credible, dated plan for those items is better than a vague yes.
Yes. The review works with what you already run: we take read-only, least-privilege access to your code, infrastructure configuration and policies under an NDA, and in total it takes between four and ten hours of your team's time, for interviews, prioritization and the readout. Most of our engagements start within two to four weeks of signature.



