Major

Vibe Coding Security: A Governance Model for Production

Vibe coding can accelerate internal software, but speed alone does not make an app safe. A practical security model ties review, permissions, data exposure, and release controls to the impact of the work.

Rahul RamakrishnanUpdated
Vibe Coding Security: A Governance Model for Production

Is vibe coding secure? Start with the app's impact

Vibe coding is neither secure nor insecure by label. The risk depends on what data the app touches, what permissions it holds, who can reach it, and what damage a wrong action can do. A prototype on synthetic data is a low-stakes experiment. An app that reads customer records and writes to a system of record is production software, whoever or whatever wrote the code.

Treat generated code as untrusted until it has been reviewed, tested and run with least privilege. The UK NCSC makes a related point in its vibe coding spectrum approach: oversight should scale with autonomy and impact. We agree, and we'd add a rule for platform teams. Never give unreviewed generated code production credentials or sensitive-data access.

Major fits at the point where this rule becomes enforceable. It is the enterprise platform where agents build the software they run on, so the thing people ship is a running app with permissions, managed state and audit logs, not a chat transcript or a folder of scripts on a laptop. That gives a Head of Platform something concrete to inspect, own and gate. It does not certify the code as secure, and it does not replace review.

This piece covers governance of the release path. It does not cover model-level attacks such as prompt injection in depth; our AI agent security threat model handles that.

Where vibe-coded apps create enterprise risk

A working demo is not security evidence. These are the failure modes that matter, none of which show up when the app runs correctly on the happy path.

  • Secrets exposure. Keys pasted into prompts, hardcoded in generated files or committed to a repository.
  • Weak access control. Authentication that works for the builder but skips authorization checks per user, role or record.
  • Unsafe dependencies. Packages the model suggests that nobody vetted, pinned or inventoried.
  • Excessive credentials. An app given an admin token because it was the fastest way to make the demo work.
  • Data handling. Sensitive records sent to unapproved tools, or logged and stored where they should not be.
  • Unclear ownership. The builder changes teams, the app keeps running, and nobody maintains it or knows it exists.

The Canadian Centre for Cyber Security covers these themes in its guidance on managing security risks in vibe coding, including sandboxing, least privilege, dependency and SBOM hygiene, human review and testing. Read it with your own app inventory open. Ownership and maintenance are the gaps most platform teams find first, because nobody scanned for them.

Risk-tier the release path, not the author

Review by impact. Whether a human or a model wrote the code is a poor predictor of risk.

  • Tier 1, synthetic-data prototype. No real data, no credentials, no external exposure. Allow freely in an isolated environment. Expire it by default.
  • Tier 2, limited internal app. Real but low-sensitivity data, a named user group, read-mostly access. Require a named owner, dependency review, secrets scanning and scoped credentials before launch.
  • Tier 3, production or high-impact app. Sensitive data or consequential actions such as writes, payments or customer messages. Require qualified human review, tests, least-privilege access, monitoring, a rollback plan and an approval recorded before release.

What to require before production

The NIST Secure Software Development Framework is the lifecycle reference here. It applies to AI-generated code the same way it applies to any code. For Tier 3, a checklist you can audit looks like this.

  1. A named owner and a backup owner.
  2. Approved tool and approved data classification recorded.
  3. Dependency list reviewed, versions pinned.
  4. Secrets scanning passed, no secrets in source or prompts.
  5. Tests run, and a qualified human has reviewed authentication and authorization paths.
  6. Credentials scoped to the minimum the app needs. See how to scope AI agent access.
  7. Logging on, with someone watching it.
  8. A tested rollback.

Platform controls improve inspectability. They do not guarantee code free of vulnerabilities.

How to build this security workflow on Major

Here is an illustrative workflow, not a customer story, for a platform team that wants every AI-built change reviewed the same way. It uses the GitHub and Slack connectors.

  • Trigger. A GitHub pull request created or updated event starts a Major workflow.
  • App that holds state. A review app stores the PR reference, data classification, owner, reviewer, configured check results, exceptions, approval status and timestamps. Because state lives in the app, a review that waits two days resumes cleanly.
  • Deterministic steps. Code in the app validates that required metadata is present and records the pass or fail results from whatever scan and test tooling you run. Those scanners are integrations you validate yourself. Major does not scan for vulnerabilities for you.
  • Agent judgment. An agent reads the diff, summarizes changes to authentication, authorization and data access, and flags unusual risks for a human. It decides what to highlight. It does not decide whether to ship.
  • Human approval. The workflow posts the review request to a restricted Slack channel and holds at an approval step. Nothing proceeds on the model's say-so.
  • Governance. The GitHub connection is read-only where possible, with any comment or status write scoped narrowly. Slack posting is limited to the review channel. The review agent never receives production credentials. Every action is logged.
  • Housekeeping. A daily schedule nudges reviewers on stale reviews, so a pending approval does not quietly become an unowned app.

The agent reasons once about how this review should run and builds the app. After that the app runs the repeatable parts the same way every time, and the model is reserved for reading the diff. See governance for agentic AI for the wider control model, and what vibe coding means if your stakeholders need the baseline.

What we're building at Major in response

Our position: AI-created software should become reviewable software that people and agents run under explicit control. Advice in a prompt or a policy PDF does not travel with the app. Controls have to live where the app runs.

That is why Major has agents build deterministic apps. The app keeps its data and history in managed storage and logs, permissions apply where an action happens, and audit records what ran and under whose access. A reviewer can inspect it, and an owner can be named for it. Reason once, run forever, with human judgment kept at the security boundary.

What we do not claim: that Major finds vulnerabilities, certifies compliance or approves releases on its own. Security engineering, code review and incident response stay with your team. Major gives them one governed place to do the work and a record of it.

Start building on Major

Build a governed review workflow for AI-generated apps, with durable approval state and scoped actions across your development process.

Related articles

Frequently asked questions

Is vibe coding secure?
It depends on the use case and the controls around it. An isolated prototype on synthetic data carries little risk. An app with sensitive data, broad permissions or consequential actions does not. Generated code should be treated as untrusted until it has been reviewed, tested and run with least privilege.
Is vibe coding a security risk?
It can be. Common risks include exposed secrets, missing authorization checks, unvetted dependencies, over-broad credentials, sensitive data sent to unapproved tools, and apps nobody owns. None of these is automatic. Risk-tiering by impact and requiring review before production keeps most generated apps manageable.
Is vibe coding safe for enterprise use?
Yes for low-impact use, with limits. Sort apps into tiers by data sensitivity and consequence. Allow synthetic-data prototypes freely, require owners and scoped credentials for internal apps, and demand qualified review, tests, monitoring and rollback for production apps handling sensitive data or consequential actions.
How to secure vibe coded apps before production?
Name an owner, record the approved tool and data class, review and pin dependencies, scan for secrets, run tests, and have a qualified human review authentication and authorization paths. Grant least-privilege credentials, turn on logging and monitoring, and confirm a tested rollback before release.
Can vibe coding tools meet enterprise security?
Tool features help, but enterprise security comes from the whole development lifecycle and organizational controls: review, testing, dependency governance, identity, monitoring and incident response. Evaluate what a tool lets you enforce and record, and do not treat any vendor claim as a substitute for your own validation.