'Plugin4Shell' Zero-Click Flaw Hit Claude Code, Codex, Copilot, Gemini CLI logo

'Plugin4Shell' Zero-Click Flaw Hit Claude Code, Codex, Copilot, Gemini CLI

AIR Security disclosed Plugin4Shell, a zero-click flaw letting a malicious plugin update bypass SHA-pinning in four major AI coding agents; Anthropic and OpenAI patched it, while Copilot and Gemini CLI remain exposed.

By the Numbers

Sept 17, 2026
Disclosed
4
AI coding agents affected
Claude Code, Codex
Patched
Copilot, Gemini CLI
Unpatched
TC
By the AI Desk
Edited by Trace Cohen · Early-stage VC & angel · Founder, New York Venture Partners
2 min read
ShareXLinkedInEmail

THE RUNDOWN

1

The flaw requires no user click, approval or reinstall -- a malicious plugin update executes attacker-controlled code purely by breaking SHA pinning, the exact mechanism developers rely on to lock an installed plugin to a specific, reviewed version.

2

It affects four of the most widely used AI coding agents at once -- Claude Code, OpenAI Codex, GitHub Copilot and Google's Gemini CLI -- via two related but technically distinct exploitation paths, one using a git branch name matching a 40-character commit hash and the other abusing Gemini CLI's FETCH_HEAD checkout behavior.

3

Compromised plugins inherit the same permissions as the developer running the agent -- local source code, cloud credentials, SSH keys, internal repositories, production systems and secrets -- meaning a successful exploit hands an attacker the same blast radius as a compromised developer laptop.

4

Patch response has split sharply by vendor: Anthropic and OpenAI shipped fixes (Claude Code 2.1.179 and Codex 0.146.0), while Microsoft has not patched Copilot and Google retired Gemini CLI without fixing the underlying flaw -- a real-world test of how differently major labs prioritize a shared vulnerability once it's public.

TC

The VC Read · Trace's Take

Trace Cohen

Two of four major AI coding agents remain exploitable right now, days after public disclosure -- that gap between patched and unpatched is the actual risk, not the vulnerability itself. The diligence item for any portfolio company running these tools in CI/CD: audit whether SHA pinning is your only plugin-trust control, because this disclosure proves that control alone doesn't hold, and a second cross-vendor plugin flaw in one year says the trust model, not any single implementation, needs rethinking.

Analysis

AIR Security researchers Or Nevo, Dor Granat and Niv Hoffman disclosed Plugin4Shell, a zero-click remote code execution vulnerability affecting Anthropic's Claude Code, OpenAI's Codex, GitHub Copilot and Google's Gemini CLI, on September 17, according to The Register and Help Net Security.

How The Exploit Actually Works

The vulnerability breaks SHA pinning, the mechanism developers use to lock an installed plugin to a specific, reviewed version of its code. For Claude Code, Codex and Copilot, an attacker can create a git branch whose name matches the plugin's 40-character commit hash; because git can prioritize a matching reference name over the underlying commit object during checkout, the agent installs the malicious branch's code while still reporting a successful installation against the expected SHA. Gemini CLI has a distinct variant of the same underlying weakness: its plugin process fetches the intended commit and then checks out a reference called FETCH_HEAD, which an attacker can hijack by naming a malicious branch FETCH_HEAD to redirect the checkout away from the actual fetched commit.

Anthropic patched Claude Code in version 2.1.179, and OpenAI patched Codex in version 0.146.0 -- both within days of the September 17 disclosure.

No Click Required, Full Developer Permissions

What makes Plugin4Shell more severe than a typical supply-chain bug is that exploitation requires nothing from the victim -- no click, no approval prompt, no manual reinstall. Once triggered, the malicious code runs with the same permissions as the developer's own agent session, which for most professional coding-agent users means access to local source code, cloud credentials, SSH keys, internal repositories, production systems and secrets. That is functionally equivalent to a compromised developer laptop, delivered through a plugin update most developers would never think to scrutinize.

A Real-Time Test Of Vendor Response

The four affected companies have responded very differently in the days since disclosure. Anthropic patched Claude Code in version 2.1.179, and OpenAI patched Codex in version 0.146.0 -- both within days of the September 17 disclosure. Microsoft has shipped no fix for GitHub Copilot as of this writing, and Google's response was to retire Gemini CLI entirely rather than patch the underlying flaw. Pulse has separately tracked a related flaw, GitSpawn, affecting overlapping coding agents earlier this year -- this is now the second cross-vendor plugin-security incident hitting the same category of tools within months, suggesting the plugin-installation trust model itself, not any single vendor's implementation, is the recurring weak point.

The Numbers In Context

A flaw hitting four of the most widely deployed AI coding agents simultaneously is a different order of severity than a single-product vulnerability -- these tools now sit inside a meaningful share of professional software development workflows, and "millions of agents affected," per AIR's own disclosure, is not hyperbole given how quickly these products have been adopted since 2024. That scale is exactly why the split patch response matters: two of the four affected products remain exploitable in production right now.

What To Watch

Whether Microsoft ships a Copilot fix, and how long Google leaves Gemini CLI users without a supported replacement that addresses the underlying checkout-validation flaw, will determine how much real-world exposure persists beyond this disclosure. For any team running these agents in CI/CD or with access to production credentials, the immediate action is auditing which plugin sources are trusted and whether SHA pinning is the only control in place -- this disclosure demonstrates that control alone is not sufficient.

ShareXLinkedInEmail

Key Sources

3 sources

Reported by The Register · First reported by AIR Security · Analysis by Value Add Pulse.

← Back to Pulse

THE WIRE in your inbox— Tech, startup & VC news with Trace's take. Free, no spam.