Skip to content
See the World Through ScienceA project of ALLATRA

Security Flaw Lets Attackers Silently Hijack AI Coding Agents via Plugin Updates

AI & Technology

Republish this story

Our work is licensed under Creative Commons BY-NC 4.0. You may republish this piece for free — with credit to ALLATRA Media and a link to the original, unedited beyond length trims, and not for commercial use.

Read the full license

A grid of GitHub Copilot product panels labeled Agent mode, MCP support, code review and next edit suggestions, with the GitHub Copilot wordmark across the center.
Product panels for GitHub Copilot's agent mode and MCP server support (illustrative). Copilot is one of four AI coding agents affected by the Plugin4Shell flaw, and Microsoft had not shipped a patch as of the Sept. 17, 2026 disclosure."GitHub Copilot: Agent mode and MCP support" by 紅色死神, via flickr, BY-NC-SA · BY-NC-SA

Air Security researchers Or Nevo, Dor Granat, and Niv Hoffman disclosed a zero-click remote code execution vulnerability on Sept. 17, 2026, affecting Claude Code, OpenAI Codex, GitHub Copilot, and Gemini CLI. The researchers named it Plugin4Shell.

The flaw sits in how each agent installs plugins from a marketplace. The standard safeguard is SHA pinning: a reviewer approves a plugin at a specific commit, and the agent is supposed to run only that exact version. Air Security found that none of the four agents verify the checkout actually resolved to the pinned commit. An attacker who controls a plugin's repository can create a branch whose name matches the pinned SHA, set it as the repo's default, and the agent's checkout resolves to the attacker's branch instead. The Gemini CLI variant works differently, exploiting a similar gap in how the CLI handles FETCH_HEAD, but the missing check is the same in both cases.

What makes it zero-click is background auto-update. Claude Code and Codex update installed plugins automatically by default, so a plugin a user already trusts can be replaced with a malicious version without any user action when the marketplace bumps the pinned version.

As of the Sept. 17 disclosure, Anthropic had patched the flaw in Claude Code 2.1.179, confirmed June 17, 2026. OpenAI confirmed a fix in Codex 0.146.0, verified Aug. 12, 2026. Microsoft had not shipped a fix for GitHub Copilot as of disclosure. Google will not patch Gemini CLI, which the company has deprecated; all installations remain vulnerable.

Air Security described the fix as an assertion that must run inside the agent after checkout: confirming the resolved commit matches the pinned SHA before execution. Because the pin is checked on the client, no marketplace change can enforce the guarantee.

Anyone running Claude Code or Codex should confirm they are on the patched versions. GitHub Copilot users have no patch available as of disclosure on Sept. 17, 2026. Gemini CLI should be replaced; no fix will be issued for the deprecated tool.

Sources

Spot an error?

Spot an error?

Report an error

Spotted a mistake on this page? Tell us what's wrong and our editors will take a look.

What kind of problem?

Only if you'd like us to be able to follow up. We won't use it for anything else.

We correct mistakes openly. Select any text to flag it. Fixes are logged under our Corrections Policy.

Report an error

Reporting on

Security Flaw Lets Attackers Silently Hijack AI Coding Agents via Plugin Updates

What kind of problem?

Only if you'd like us to be able to follow up. We won't use it for anything else.

We read every report. Corrections are logged publicly.