Ryz Labs/SOC 2 for AI-assisted code
Ryz Labs

SOC 2 and security reviews for AI-assisted code

What auditors and enterprise security reviewers check in codebases built with Cursor, Copilot, Claude Code, Lovable or Replit, mapped to the trust criteria, plus a 15-point readiness checker.

If an enterprise deal is waiting on your security review, or your SOC 2 audit window is coming up, the code written with AI assistants is held to the same standard as every other change. The trust services criteria ask whether changes were authorized, tested and approved, whether access is limited, and whether vulnerabilities are found and fixed. They do not care who or what typed the code. This guide maps the gaps we see most in AI-assisted codebases to the criteria and to the tests compliance platforms run, then shows how engineers on your team close them.

What this page is not. A SOC 2 report is an attestation issued by a licensed CPA firm under AICPA standards (AICPA SOC suite). Compliance platforms collect evidence, and engineers fix code and process. Ryz Labs does not audit, attest or certify anything. Our engineers work on your team, reporting to your leads, to get your repos, pipelines and data model ready for the people who do.

15-point AI-code SOC 2 readiness checker

Answer each check honestly. If you are not sure, answer No: an unknown is a gap until someone has checked it. The checker runs in your browser and stores nothing.

Code review and change provenance
01
Every change to the production branch goes through a pull request approved by someone other than the author.Vanta and Drata both test for this on GitHub. A branch where one person can merge their own change fails the review test. CC8.1
02
New commits dismiss earlier approvals, and admins cannot bypass the branch rules.Otherwise a change can be approved, then rewritten by an assistant, then merged with the old approval. CC8.1
03
AI-assisted changes get the same review, tests and checks as any other change, and each pull request shows who made it and who approved it.Auditors test the process, not the tool. A commit pushed straight from an AI builder or agent with no pull request is the gap they find. CC8.1
04
Only people in your identity provider can write to production code, and that access is reviewed.Former contractors, personal accounts and shared bot tokens with write access are common findings. CC6.3
Secrets
05
No secrets live in frontend code, client bundles, the repo or its git history.Scan the full history, not just the latest commit. Anything that was ever committed should be treated as exposed. CC6.1, CC6.7
06
Push protection or a pre-commit scan blocks new secrets before they reach the repo.AI assistants paste keys into config files and examples. Blocking them at push time is cheaper than rotating them later. CC6.8
07
Service-role and other admin keys are used only on the server, and any exposed key has been rotated.On Supabase the secret key bypasses every row-level security policy, so one leaked key exposes every table. CC6.1
Data isolation
08
Every table and endpoint holding customer data enforces ownership or tenant on the server, not just "logged in".Row-level security policies or server-side checks. Hiding a button in the UI is not access control. CC6.1, CC6.3
09
Automated tests try to read and write another tenant's data, and fail.Security reviewers and pen testers try exactly this. Tests keep it from coming back after the next generated change. CC6.1
Dependencies and scanning
10
Dependency alerts and automated security updates are on, and new packages are checked before they are added.Code assistants sometimes suggest packages that do not exist, which attackers can register under that name. CC7.1, CC9.2
11
Static analysis (code scanning) runs on every pull request as a required check.It catches injection and unsafe patterns that reviewers skim past in large generated diffs. CC7.1
Change management and CI/CD
12
Tests, lint and scans must pass before merge, and production deploys come only from the pipeline.No deploys from a laptop or straight from a builder. The pipeline is your evidence that changes were tested before release. CC8.1
13
Emergency changes have a written path, with review after the fact.The criteria expect a process for urgent changes too. Hotfixes are where review usually gets skipped. CC8.1
Operations
14
Vulnerabilities have fix deadlines by severity, and you can show they are met.Compliance platforms track findings against deadlines. Open critical findings fail their tests until they are fixed. CC7.1
15
Logs and alerts cover sign-ins, admin actions and production changes, and are kept for your audit period.A Type 2 report covers a period of time, so the evidence has to exist for the whole period. CC7.2
Context (not scored)
+
Does your product call an LLM API?Adds an LLM feature security workstream based on the OWASP Top 10 for LLM Applications.
+
Do you handle health data (PHI) for healthcare customers?Adds a BAA and HIPAA configuration workstream.
Answer the checks 0 of 15 answered

Your prioritized remediation tickets appear here as you answer.

    This is a readiness self-check, not an audit opinion. Only a licensed CPA firm issues a SOC 2 report.

    Scope your team →

    Why AI-assisted code shows up in security reviews

    Peer-reviewed studies found the gap years ago. In a Stanford-led study published at ACM CCS 2023, participants who used an AI assistant wrote significantly less secure code than those who did not; the OpenSSF summarizes one result as 36% of assisted participants writing solutions vulnerable to SQL injection, against 7% of the control group (OpenSSF guide). An IEEE S&P 2022 study of GitHub Copilot found about 40% of 1,689 generated programs were vulnerable. A 2025 vendor report from Veracode says 45% of AI-generated code samples failed its security tests.

    The tool makers say the same thing. GitHub's Copilot guidance says "Review Copilot's suggestions carefully" and recommends automated tests, linting and code scanning to check its work. The OpenSSF guide for AI code assistants puts it plainly: AI-generated code is not a shortcut around code review and testing.

    None of this means AI-assisted teams fail audits. It means the volume of change goes up while review stays human, and the controls that prove review happened become the bottleneck. That is where enterprise questionnaires and auditors look.

    What SOC 2 actually checks

    SOC 2 reports cover up to five categories: security, availability, processing integrity, confidentiality and privacy. The common criteria are what evaluate security, so security is in every report (AICPA TSC). A Type 1 report describes your controls at a point in time. A Type 2 report adds the auditor's opinion on whether those controls operated effectively throughout a period, with a description of the tests the auditor ran. The criteria also state that a company does not have to address every point of focus; they are guidance, not a checklist.

    For engineering teams, four criteria do most of the work:

    Why now. On October 28, 2026 the AICPA proposed an update to the SOC 2 criteria with new points of focus on AI governance, model monitoring, prompt manipulation, data poisoning and the transparency of AI-generated outputs. Comments are due January 15, 2027, and if approved the update would apply to engagements beginning on or after December 31, 2027, with early adoption permitted. Teams that clean up AI-era change control now will not have to do it twice.

    The gaps we fix most, mapped to the criteria

    Pull request review and change provenance (CC8.1, CC6.3)

    This is the gap compliance platforms test most directly. Vanta's Application Changes Reviewed test for GitHub requires a pull request before merging with at least one approval on the production branch. Drata's Test 8 fails if a user can merge their own change with no one else involved, its Test 9 checks that only authorized people can change production code (rulesets left in Evaluate mode do not count), and Test 7 cross-checks repository write access against your identity provider.

    AI-assisted work breaks these in predictable ways: an agent or builder pushing straight to main, one engineer approving a large generated change and then pushing more commits on top, or admins bypassing rules to unblock a release. GitHub's branch protection can dismiss stale approvals when new commits change the diff, require that the most recent push be approved by someone other than the person who pushed it, require code owner review, and apply the rules to admins too. CODEOWNERS routes changes to sensitive paths to the right reviewers automatically.

    No Vanta or Drata test we found targets AI-written code specifically: the tests check that a change was reviewed and approved. So the fix is process, not tooling. Route every builder and agent change through a pull request, label AI-assisted work the way your team agrees, and make the reviewer accountable for the diff.

    Secrets in code, bundles and history (CC6.1, CC6.7, CC6.8)

    Lovable's own security guidance says secrets stored in frontend code should be considered compromised. On Supabase, the secret key bypasses every row-level security policy, and Supabase is deprecating the legacy anon and service_role keys by the end of 2026, which is a natural moment to rotate.

    GitHub push protection blocks pushes that contain secrets before they reach the repository; for repositories it requires GitHub Secret Protection and is off by default. Secret scanning runs free on public repositories, and private repositories in an organization need Secret Protection. By default, anyone with write access can bypass push protection with a reason, and each bypass creates an alert and an audit log entry, which is the evidence your reviewer will ask about.

    Row-level security and tenant isolation (CC6.1, CC6.3)

    Supabase is explicit: enable RLS on every table in an exposed schema, because a table without it is readable and writable by any role with a grant. Views bypass RLS by default unless they are created with security_invoker. The NVD record CVE-2025-48757 describes insufficient row-level security in sites generated by Lovable before April 15, 2025; Lovable disputes the CVE, saying each customer is responsible for protecting its application's data. Either way, the reviewer's first test is the same: log in as one customer and try to read another customer's rows.

    Policies that check only "is logged in" pass a demo and fail that test. Fix them table by table with ownership or tenant checks, close any view or function that skips the policies, and add cross-tenant tests to CI so the next generated migration cannot reopen the hole. The OWASP Top 10 lists Broken Access Control first for a reason.

    Dependencies and supply chain (CC7.1, CC6.8, CC9.2)

    Code assistants add dependencies fast, and some suggested packages do not exist: the OpenSSF guide cites research finding about 19.7% of packages suggested by code models were not real, which attackers can register. GitHub Dependabot alerts are based on the GitHub Advisory Database, and security updates open pull requests to the minimum patched version. Code scanning with CodeQL automates security checks on every change; on private repositories it needs a GitHub Code Security license.

    For build provenance, GitHub artifact attestations are signed claims about how an artifact was built and on their own provide SLSA v1.0 Build Level 2. SLSA defines provenance as what built the artifact, the process used, and the inputs. Enterprise reviewers increasingly ask for it, and it answers "where did this binary come from" without anyone digging through chat logs.

    CI/CD gates and change management (CC8.1)

    Required status checks mean tests, lint and scans must pass before a change merges (GitHub docs). Signed commits can be required too. The pipeline becomes your evidence for CC8.1: who authorized the change, which checks ran, who approved it and when it shipped. GitHub's organization audit log keeps 180 days of events and only owners can read it, so export what your audit period needs.

    Two gaps show up repeatedly in AI-assisted teams: deploys that skip the pipeline (from a laptop or straight from a builder), and hotfixes with no review afterwards. The criteria expect an emergency-change process, so write one down and use it.

    Vulnerability deadlines and pen-test findings (CC7.1, CC4.1)

    When no policy is configured, Drata's defaults for fixing vulnerabilities are 7 days for critical, 30 for high, 90 for medium and 180 for low, and its Critical and High Vulnerabilities Addressed tests fail while any finding of that severity is open (Drata docs). Vanta tracks remediation against the deadlines you configure (Vanta GitHub guide). CVSS v4.0 rates 9.0 to 10.0 as critical and 7.0 to 8.9 as high (FIRST).

    The trust criteria mention penetration testing only as one example of a monitoring evaluation, and do not require every point of focus, but enterprise buyers often ask for a recent pen test anyway. When the report lands, the work is fixing the findings in priority order and showing the evidence, which is engineering work, not paperwork.

    Runtimes and databases past end of life stop receiving security fixes, so their findings never close. The free EOL deadline radar reads a package.json, requirements.txt, Gemfile or Dockerfile and lists every dated deadline in your stack, with links to the official calendars.

    LLM features in the product (OWASP LLM Top 10)

    If your product calls a model, reviewers increasingly ask how you handle the OWASP Top 10 for LLM Applications 2025: prompt injection, sensitive information disclosure, supply chain, data and model poisoning, improper output handling, excessive agency, system prompt leakage, vector and embedding weaknesses, misinformation and unbounded consumption. NIST's SP 800-218A adds secure development practices for generative AI on top of the SSDF. Expect the proposed 2027 SOC 2 criteria to bring prompt manipulation and AI output transparency into audits directly.

    Health data and BAAs (HIPAA)

    HIPAA treats anyone who creates, receives, maintains or transmits protected health information for a covered entity as a business associate, including subcontractors (45 CFR 160.103). For an AI-built health app, that usually means a BAA with each vendor in the data path. Supabase offers HIPAA as a paid add-on on Team and Enterprise and requires a signed BAA (Supabase docs); its own SOC 2 report does not transfer to your environment. OpenAI lets API customers accept a standard BAA, Anthropic provides a BAA for HIPAA-ready services with listed exclusions, and AWS lists Amazon Bedrock as a HIPAA eligible service under its BAA. Check the covered services for each vendor before PHI flows through them.

    One-page map

    Use this to brief your auditor, your compliance platform owner and the engineers doing the work.

    GapCriteriaWhat gets testedThe engineering fix
    Pull request reviewCC8.1Vanta Application Changes Reviewed; Drata Tests 8 and 9Required approvals, stale-approval dismissal, last-push approval, CODEOWNERS, no admin bypass
    Who can change codeCC6.3Drata Test 7Write access reconciled with your identity provider, scoped tokens
    SecretsCC6.1, CC6.7, CC6.8Secret scanning alerts and bypass recordsHistory scan, rotation, server-side keys, push protection
    Data isolationCC6.1, CC6.3Pen test and questionnaire findingsRLS or server-side checks on every table, cross-tenant tests
    DependenciesCC7.1, CC9.2Dependabot and scanner findings by severityAlerts and updates on, package verification, code scanning
    Build and releaseCC8.1Automated checks on code changesRequired status checks, pipeline-only deploys, emergency path
    Vulnerability deadlinesCC7.1Critical and high findings testsSeverity deadlines, triage, fixes with evidence
    MonitoringCC7.2Log retention and alerting evidenceAuth, admin and deploy logs kept for the audit period

    Criteria text: AICPA TSC (2017, March 2020 update; the current edition revises points of focus). Platform tests: Vanta and Drata help centers. GitHub settings: GitHub docs. All checked November 21, 2026.

    How engineers on your team close the gaps

    Phase 0: Baseline

    Export branch rules, repository access, secret scanning and dependency findings, and your compliance platform's failing tests. Agree with your auditor or the customer's reviewer what evidence they expect and for which period. Exit: one list of open gaps with owners.

    Phase 1: Blockers

    Branch protection and review rules, secret rotation and server-side keys, row-level security and server-side authorization on customer data, pipeline-only deploys. Exit: the platform's review, access and critical-vulnerability tests pass, and a cross-tenant test runs in CI.

    Phase 2: Evidence that runs itself

    Required status checks for tests, lint and scans; dependency alerts and code scanning wired to the compliance platform; artifact attestations; log retention for the audit period. Exit: evidence updates without anyone taking screenshots.

    Phase 3: Process for AI-assisted work

    How AI-assisted changes are labeled and reviewed, assistant configuration and rules files under version control, package verification, the emergency-change path, and the LLM feature threat model if you call a model. Exit: written procedures that match what the pipeline enforces.

    What an AI pod does in the first two weeks

    These are activities, not promises about your audit. Our AI pod uses coding assistants for the slow parts (policy drafts, test generation, scanner triage), and every change still goes through human review and your pipeline.

    What it costs

    Ryz engineers on your team start from $7,000 per month. Our published range is $7,000 to $15,000 per engineer per month, mostly mid-level to senior, quoted per team. Many teams begin with a pod of 3 to 5 senior engineers (a tech lead plus engineers), and a full pod is about seven; a focused remediation can start with a tech lead plus one or two engineers. Once you say go, engineers can typically start in under two weeks. See engineer rates or scope your team.

    Related

    FAQ

    Does AI-generated code fail SOC 2?

    Not by itself. The trust services criteria ask whether changes were authorized, tested, approved and implemented under control, and whether access and vulnerabilities are managed. They do not depend on who or what wrote the code. AI-assisted teams run into trouble when changes skip pull requests, approvals go stale, secrets land in code, or data isolation is weak.

    Can Ryz Labs certify us for SOC 2?

    No. A SOC 2 report is issued by a licensed CPA firm under AICPA standards, and compliance platforms such as Vanta and Drata collect evidence. Ryz engineers work on your team to fix the code, pipeline and process gaps that auditors and customer security reviews find.

    What do Vanta and Drata check in GitHub?

    Vanta's Application Changes Reviewed test requires a pull request with at least one approval on the production branch. Drata's tests check that no one can merge their own change alone, that only authorized people can change production code, and that repository write access matches your identity provider. Both track vulnerability findings against fix deadlines.

    Is a penetration test required for SOC 2?

    The trust services criteria mention penetration testing as one example of a monitoring evaluation and state that not every point of focus has to be addressed, so the criteria do not require one. Many enterprise customers ask for a recent pen test in their security review anyway, and fixing the findings is engineering work.

    How long does it take to fix SOC 2 gaps in an AI-assisted codebase?

    It depends on how many gaps are open, how much customer data is live, whether row-level security needs rework, and how your audit period or customer deadline falls. We scope it with you after reading the code and your platform's failing tests, instead of quoting a number before anyone has looked.

    Thanks — your message has been sent. We’ll get back to you soon.

    Something went wrong while sending your message. Please try again or email info@ryzlabs.com.

    Explore Ryz Labs

    Software development servicesAI development companyIndustriesStaff augmentationAI pod teamsNearshore software developmentHire engineers by roleEngineer ratesGuidesBuyer guidesCase studiesHow we vet engineers

    Ryz Labs

    Senior engineers on your team. AI pod teams that ship.

    Tell us what you're building. You get a scoped plan, a price and the names of the people who would do the work.

    • Only the top 1% of tens of thousands interviewed make it
    • On US business hours, including New York hours
    • Trusted by Fortune 500 engineering teams

    Tell us who you need

    A Ryz partner replies with a scoped team plan.

    Start a conversation →

    Prefer to talk? Book a 15-min call →