Vibe Coding and Supply Chain Security:
The Unreviewed Dependency Problem.

Dependency Churn in AI-Assisted Projects

Traditional projects acquire dependencies deliberately: a developer identifies a need, evaluates options, chooses a package, and adds it to the dependency manifest. Each addition is a conscious decision with at least some evaluation of the package.

In vibe-coded projects, dependencies arrive as outputs of code generation. An AI that generates a feature will also generate the import statements and sometimes directly modify the package manifest. The developer may not register which new packages were added as a side effect of asking the AI to "add file upload support" or "implement rate limiting."

Where the Trust Model Breaks

Traditional supply chain security is built on the assumption that a developer chose each dependency. That assumption gives us useful heuristics: a senior developer who has used a package for years is unlikely to add a malicious version of it by accident.

Vibe coding breaks this assumption in several ways:

• No package familiarity — developers using AI-generated code may have no prior knowledge of the packages being added to their project
• No evaluation step — the decision flow goes from "tell AI what I want" to "merge the PR" without a distinct evaluation step for new dependencies
• Context collapse — a developer reviewing a 500-line AI-generated PR may not notice three new entries in package.json that weren't there before
• Velocity pressure — the value proposition of vibe coding is speed; pausing to evaluate each new dependency slows down the workflow that makes AI coding attractive

Agentic Tools and Autonomous Installs

Agentic coding tools that can execute shell commands create a critical escalation: the AI can install packages directly, without any human approval step. An agent tasked with "implement this feature" may run npm install or pip install as part of its implementation — potentially adding packages that were never reviewed by a human.

User: "Add email validation to the signup form"

Agent actions:
1. Reads current form code
2. Decides email-validator package would help
3. Runs: npm install email-validator-extended  ← not reviewed
4. Writes new form code using the installed package
5. Reports: "Done! I added email validation using email-validator-extended"

# email-validator-extended doesn't exist officially
# An attacker registered it with a credential-stealing payload
# The agent just installed malware autonomously

Transitive Dependency Explosion

AI coding tools are not sensitive to the difference between a package with 5 transitive dependencies and one with 500. They optimise for feature completeness and API ergonomics — not minimal dependency footprint.

A vibe-coded project that uses AI-suggested libraries may end up with significantly deeper and wider dependency trees than a project where a developer made explicit choices about dependency weight. Each transitive dependency is supply chain exposure — a potential CVE, a potential compromise, a package that could be typosquatted at a higher version.

Security teams scanning AI-assisted projects often find the dependency count has grown dramatically compared to similar traditionally-coded projects, increasing the continuous monitoring burden proportionally.

SCA Gaps in Vibe Coding Workflows

Traditional SCA tooling was designed for a world where dependency changes are deliberate and infrequent. In vibe coding workflows:

• Alert fatigue — high dependency churn generates more SCA alerts; security teams may start tuning out or auto-dismissing alerts that arrive too frequently
• Lock file verification gaps — if agentic tools can modify lock files without going through PR review, lock file integrity is compromised
• SBOM staleness — SBOMs generated at release time may not reflect the current dependency state of rapidly iterated vibe-coded projects
• Provenance tracking gaps — knowing why a dependency was added (which developer, which PR, for what reason) becomes harder when AI agents add them autonomously

Adapted Security Controls for Vibe Coding Projects

• Block agent package installs — configure agentic tools so they can suggest package additions but cannot install them directly; all installs require human approval
• New dependency alert on every PR — automatically flag any PR that adds new packages (direct or transitive) for security review, with metadata on the package's age, download count, and provenance
• Dependency allowlist enforcement — AI-assisted projects benefit most from strict allowlisting; packages not on the list cannot be added without security team approval
• Continuous SCA with real-time alerts — don't wait for CI; scan for new CVEs in your dependency tree continuously and alert within hours of disclosure
• Minimal footprint prompting — include instructions in AI system prompts to prefer packages already in use over new dependencies, and to avoid packages with large transitive dependency trees