Key Takeaways
- LLMs make code generation cheap, tempting teams to rewrite third-party packages in-house to dodge supply chain attacks, but this trades upstream security alerts for permanent internal maintenance.
- Jay Parikh warns that replacing open source packages leaves your team on the hook for patches: “I have 2,000 open source projects I use in my thing. I'm going to get rid of, you know, 1980 of them and then 20 of them I'll use, but the other 1980 I've got to maintain now, right?” (play from 18:19)
- GitHub is countering supply chain risk through tools like Code Review Max to catch high-severity, obscure vulnerabilities in code before deployment.
- In a recent keynote demonstration, Microsoft's Damian accidentally pushed an active secret, and GitHub Advanced Security blocked the commit from entering the repository in real time.
- Microsoft and GitHub security teams are coordinating directly with external maintainers to lock down public package registries like npm rather than retreating behind closed walls.
The Disagreement
A growing camp of software teams argues that LLMs have changed dependency economics. Because generative models produce functional code in seconds, founders wonder why they should accept the risk of poisoned npm dependencies, typosquatting, or compromised maintainer accounts. The proposed fix sounds simple: purge third-party packages, use AI agents to generate in-house replacements, and vendor the results into private repositories.
Jay Parikh sees this as an expensive mistake. Replacing thousands of open source packages with private code isolates a codebase, but it kills the network effect that keeps software alive. In Parikh's words: “I think that there's just so much happening that's important and valuable in open source. And I don't think that's going to stop. If anything, I think the value of open source to businesses, to builders, to creators out there only increases from here. And again, I think let's solve the problem, which is like let's make the stuff in open source secure versus everybody going into like some walled garden approach.” (play from 18:39)
Retreating into a walled garden converts community-maintained logic into unmaintained private technical debt. An LLM can write a markdown parser or a cryptographic utility in twenty seconds. It will not proactively patch zero-day exploits, track evolving standards, or fix obscure edge cases across operating systems when your engineers are sleeping.
Who's Right (and When They're Wrong)
Parikh is right for 95% of a typical modern stack. If your application pulls in standard utilities, database drivers, or web framework components, writing private versions is suicide by maintenance. You forfeit updates from thousands of contributors who patch corner cases for free.
GitHub is betting that platform-level defenses beat retreat. During a recent product keynote, presenter Damian attempted to push a credential directly into a repository. GitHub Advanced Security caught the secret at the pre-receive hook and blocked the push from ever touching the repo. Combined with tools like Code Review Max that hunt for high-severity vulnerabilities during pull requests, automated enforcement handles malicious behavior at the platform layer.
The rewrite camp is right only in two narrow windows: single-function micro-packages and isolated security perimeters. If your dependency tree pulls in trivial packages with six lines of logic that introduce transitive risk, vendor those utilities directly. The same applies to air-gapped systems under strict compliance regimes where external network access is prohibited. Outside those limits, building private clones of 1,980 dependencies burns engineering cycles on undifferentiated maintenance.
What to Do With This
Run an audit on your lockfiles this week to find your total dependency count. Separate trivial single-function utilities from complex packages. Vendor the three simplest utilities into your codebase, delete their external packages, and turn on proactive secret scanning and automated pull request reviews for everything else.