Skip to main content
Back

Claude Code's Silent Dependency Choices Put Security Teams on Alert

A new pre-registered study protocol examines whether configuring Claude Code and OpenAI Codex can actually control which libraries they silently pull into your codebase.

Paper, Full Access
  • Core finding: A pre-registered study protocol is the first controlled experiment designed to test whether configuring Claude Code and OpenAI Codex can govern their silent library import decisions.
  • The problem: Agentic coding tools routinely add external dependencies without announcing them in natural language, creating hidden security and licensing risks.
  • What's being tested: Six configuration levels are compared, from no config to context files, Skills (SKILL.md), MCP-connected package registries, and permission controls.
  • Three outcomes measured: Which library the tool picks, whether it discloses the new dependency, and whether that disclosure is accurate and complete.
  • Prior evidence: Research found LLMs unnecessarily overuse popular libraries like NumPy in ~45% of cases, suggesting systematic bias that configuration may or may not override.
  • Artifact release: The benchmark dataset, config artifacts, and analysis pipeline will be released as a reusable tool for evaluating future agentic coding tools.

Every time you hand a task to an agentic coding tool, it makes a quiet judgment call: should it write this functionality from scratch, or just import something? That decision, which researchers call the build-versus-buy choice, is happening thousands of times a day across codebases worldwide. And until now, nobody has run a controlled experiment to figure out whether developers can actually influence it.

A new pre-registered study protocol from researchers at the University of Canterbury and King's College London aims to change that. The paper lays out a rigorous experimental design to test how different configuration mechanisms in Claude Code and OpenAI Codex shape the dependency decisions those tools make, and whether the tools honestly disclose the libraries they introduce.

The Silent Dependency Problem

Agentic AI coding tools accept high-level task descriptions and plan, write, and execute code with minimal per-step human direction, introducing external dependencies without explicit instruction. For example, when a developer asks such a tool to implement authentication, it may write import bcrypt or issue a pip3 install passlib command without stating so in its natural-language reply. The dependency enters the codebase unannounced.

If the developer does not review code changes systematically, it may not be noticed until a security audit or license scan surfaces it. That is not a hypothetical edge case. As adoption of these tools grows, observational data confirms that systematic library preferences are already visible in deployed releases. The scale of the problem compounds quickly: prior research found that eight LLMs systematically overuse popular libraries like NumPy unnecessarily in roughly 45% of cases, and defaulted to Python even for performance-critical tasks where it is suboptimal.

The agentic AI setting inherits traditional build-vs-buy tensions and adds two of its own: the decider is the tool rather than a human engineer, and the decision is often unannounced in the natural-language interface, so the human oversight that earlier software engineering literature relied on is harder to apply.

Why Configuration Is the Lever

Both Claude Code and OpenAI Codex expose a layered set of mechanisms that developers can use to shape tool behavior. Understanding these layers is key to understanding what the study is actually testing: