Vibe coding adds dependencies faster than human review or a point-in-time SBOM can track. An inventory produced for an audit is stale before the next release, which leaves the pipeline enforcing yesterday’s picture of the software. Treating SBOMs as living assets and checking them against policy gates at commit, build, deploy, and runtime is what keeps compliance and security in step with AI-assisted development.
How vibe coding has changed software development
Software development has changed more in the last two years than in the decade before it. AI coding assistants began as autocomplete tools that finished a line or suggested a function to now write substantial portions of production code, often from a short natural language prompt. The Stack Overflow 2025 Developer Survey found that 84% of developers use or plan to use AI tools, up from 76% in 2024. DORA 2025 reported that 90% of technology professionals use AI at work. The pace makes sense given what changed for the developer – writing code became easier, carried less cognitive load, and produced a strong feeling of productivity. Developers could describe an outcome and watch working code appear, so the habit spread without much pressure from management. One consequence is that commits and releases now arrive far more often than they did before. More code is shipping faster, but the systems it runs in are not more secure. The Veracode 2025 GenAI Code Security Report found that AI introduces security vulnerabilities in 45% of coding tasks. No team can realistically human-review every change at that volume, which means the controls that depended on review need a second look.
What security risks does AI-generated code introduce?
Though AI coding agents are trusted to generate code, they do not have awareness of your architecture, your internal stack, or the conventions already present in your codebase. They work from the prompt context and from patterns learned in public training data, which doesn’t fit your environment. Agents also have no awareness of organizational security or compliance policies. For example, they do not check a package against known CVEs before writing the import statement, so a vulnerable or banned library can enter a project in the first commit. Agents also favor pulling in a ready-made package over writing a small piece of custom logic. That habit adds new dependencies, including orphaned, duplicated, or unmaintained packages, and the unclear provenance behind some of them can cause license drift.
Package hallucination adds a supply chain problem of its own. LLMs sometimes suggest dependency names that do not exist, and a study reported by BleepingComputer found that nearly 20% of AI-suggested dependencies were hallucinated package names. Attackers can register those names and wait for a developer or agent to install them. The attack surface now reaches the developer environment itself. Prompt injection can steer an agent into running malicious instructions, and LLMjacking lets attackers abuse stolen access to model services. Both extend risk beyond the code and into the tools that produce it.
Why do point-in-time SBOMs go stale?
Vibe coding has greatly accelerated release frequency along with output. DORA’s 2025 State of AI-assisted Software Development report found that higher AI adoption is associated with an increase in both software delivery throughput and delivery instability. As a result, GitClear’s research shows code churn (code rewritten or deleted soon after creation) and small, frequent commits have made development move at a pace that’s unmanageable. Each of those changes can add or swap a dependency that no one reviewed closely.
Teams have relied on the software bill of materials, or SBOM, to record what a build contains, and it has always been a snapshot taken at a single point in time. However, a point-in-time SBOM cannot keep up with the pace of vibe coding. An inventory generated for a quarterly audit or an annual release goes stale within hours, sometimes minutes, once developers ship several times a day with AI assistance. The document ends up describing software that no longer runs in production.
An SBOM filed for reporting or a compliance checklist answers a question that was true at that point in time. Put the same inventory to work as the facts the policy gates check, and block builds that introduce a prohibited license or a known vulnerable package before they ship. That change in role turns an SBOM from a document into an asset, and it depends on preserving each SBOM for versioning purposes. When an issue surfaces, the stored history shows when it entered the codebase and what the software looked like at that moment.
Johnny Halife, CTO and Partner at SouthWorks, observed that when organizations rushed to adopt AI coding agents their “processes couldn’t keep up with the increased velocity.” They generated more code than they could properly review or validate. The existing processes in organizations were designed for a slower pace, and they fail once output multiplies.
At AI scale, human review and point-in-time checks cannot carry the load. SBOMs answer this problem because they turn the contents of your software into data that a pipeline can check automatically on every commit.
Why an SBOM is an asset, not an artifact
Kept as a living inventory rather than a one-time filing, an SBOM stays accurate for each build and can be queried whenever a new vulnerability is disclosed. SBOM-based policy gates belong at four points in the pipeline. At commit, an SBOM is generated carrying a record of every change introduced. At build, that record is scanned for vulnerable or unwanted packages. At deploy, policies are applied to decide whether the artifact may reach production, and the SBOM acts as attestation for any image or application that makes it that far. At runtime, stored SBOMs are assessed again as new vulnerabilities are disclosed.
Continuous runtime checks are what give the stored SBOMs their value. Each SBOM records what a build contained at a point in time, so a new disclosure can be measured against the full history rather than the single artifact running today. When a vulnerability surfaces months after release, the team can answer from the SBOM-based data it already holds instead of rescanning every artifact by hand.
The two approaches differ in when the inventory is created and what a team can do with it.
| Point-in-time SBOM | Continuously enforced SBOM | |
| Generated | Once, for an audit or a release | At commit, build, deploy, and runtime |
| Describes | The software as it was at that moment | The software as it changes |
| Freshness | Stale within hours of the next commit | Current with every change |
| Role | Evidence filed for a compliance checklist | Evidence used by policy checks at every gate |
| On a new disclosure | Each artifact must be rescanned by hand | The stored inventory can be queried |
| History | Not retained | Shows when an issue entered the codebase |
How do you get started with SBOMs for AI-generated code?
Getting started comes down to four choices about removing friction rather than adding process:
- Remove friction for developers: Adopting SBOMs works best when the process stays out of the developer’s way. Generation, scanning, and enforcement need to run continuously and fast, so they fit inside the normal workflow instead of becoming a roadblock developers wait on. Every decision the pipeline makes should come with a clear explanation. A blocked build should tell the developer what triggered it and what to change.
- Start narrow: Start by picking the projects that lean most heavily on AI-generated code, since those see the highest dependency churn and show returns the fastest. A small pilot lets the team tune the pipeline before it reaches every repository, and it produces early evidence that helps justify wider rollout.
- Block sparingly: Define three blocking policies to begin with. Keep the first set small and enforceable, covering the failures the team cares about most. A policy list that grows too long ends up with nothing that actually blocks, because every build trips something and exceptions pile up. Add more policies later, once the first three are trusted and the team understands how they behave in practice.
- Centralize the inventory: By storing all SBOMs in one place, a new vulnerability can be looked up against everything that has shipped rather than one artifact at a time. That turns a disclosure from a manual search across repositories into a single query, so the team can see which releases are affected without chasing each one.
SBOMs are no longer just compliance artifacts. Treated as a living inventory, they let teams move at AI speed while staying secure and compliant. The inventory tracks what ships with each change, so the checks keep pace with the commits.
Key takeaways
- AI coding agents introduce dependencies faster than human review can absorb.
- A point-in-time SBOM goes stale within hours, so it stops describing what is running now.
- An SBOM works better as a living inventory with policy gates than as an artifact filed for an audit.
- Policy gates belong at commit, build, deploy, and runtime, with relevant SBOMs at each stage stored centrally so history stays traceable.
- Continuous scanning of stored SBOMs answers new disclosures without rescanning each artifact by hand.
- Getting started takes one project, a few blocking policies, and one place to store the inventories.
If your team has not started, begin with a single project. Generate an SBOM, scan it, write a few blocking policies, and test whether the process holds up under real commits. Anchore Enterprise stores those SBOMs and applies policy across the organization, while AnchoreCTL, its command-line client, is how the checks get wired into the pipeline. Anchore Enterprise also provides visibility into the AI models in use and how they generate code. An SBOM records components rather than authorship, which allows you to distinguish code a developer wrote and reviewed by hand from that generated by an agent.
Vibe coding will keep changing how software gets written, and the teams that thrive with it will be the ones whose SBOMs work as living infrastructure rather than a yearly filing.