how it happened
- whoanyone with a GitHub accountcould open an issue on the public repository
- the gapan AI triage bot with a shellthe title went into its prompt, Bash was allowed
- thenthe nightly release workflowit held the npm and extension publish tokens
- the resultcline@2.3.0 on npmit installed another AI agent for about eight hours
what everyone assumed
Cline is an open-source AI coding assistant, shipped as editor extensions and as a command-line tool on npm. As an open-source project, it receives GitHub issues from strangers. On 21 December 2025, according to Cline’s post-mortem of 24 February 2026, the project added a workflow that let an AI agent triage them: each new issue started Claude, through Anthropic’s claude-code-action, to read it and respond.
Cline has not published what its team believed about that bot, so this belief is an inference from how it was set up. A triage bot looks harmless. Its job is labels and replies, and the post-mortem says the workflow had limited GitHub permissions: read-only access to the code and write access to issues. Nothing in that list touches a release. That reading was reasonable, and it missed two things the permissions list did not show: what the agent could run, and what the machine it ran on could touch.
what happened
The bot read text from anyone and had a shell. Adnan Khan, the researcher who found the flaw, published the workflow’s settings on 9 February 2026. Any GitHub account could trigger it, the issue title was placed directly into Claude’s prompt, and the allowed tools included Bash, file writing and web fetching. Anthropic’s security documentation for the action, as it read on 13 November 2026, says that by default only users with write access can trigger it, and calls the setting that opens it to everyone “a significant security risk”. Cline’s post-mortem sums up the combination as “functionally equivalent to giving every GitHub user shell access”.
The shell reached the release through the cache. Khan showed that a crafted issue title could make Claude install a package from an attacker’s commit, which runs that commit’s code on the build machine. He writes that the triage bot’s own GitHub token was restricted and that its only secret was the Anthropic API key. The way onward was the build cache. Cline’s post-mortem confirms that the triage workflow shared its cache scope with the nightly release workflow. Khan describes filling the cache with more than 10 GB of junk to force old entries out, then writing poisoned entries under the names the nightly workflow expected. That workflow held the publish tokens for the VS Code Marketplace, OpenVSX and npm. Khan says he ran his proof of concept on a mirror of the repository with his own API key, without testing against Cline itself.
The report waited. Khan’s timeline says he reported through a GitHub security advisory on 1 January 2026 and followed up on 8 January, 18 January and 7 February, the last time receiving only a ticket number. He published on 9 February. By Cline’s account it removed the vulnerable workflows within about half an hour and began rotating credentials.
The rotation missed one token. The post-mortem says the extension tokens were rotated, but for npm the wrong token was deleted while the exposed one remained active. The team checked its work on npm’s organisation token page, which showed zero active tokens and did not list the exposed one.
A release went out. GitHub advisory GHSA-9ppg-jx86-fqw7, published by Cline on 17 February 2026, says an unauthorised party used a compromised npm publish token to publish cline@2.3.0 at 03:26 Pacific time. The only change was an install script that installed OpenClaw, a separate open-source AI agent, globally on the machine. The rest of the package, including the program itself, matched version 2.2.3, and the editor extensions were not affected.
Who published it, and how they got the token, is not established in any public source. Cline calls it “the exposed” token in the same account as the triage flaw, without saying in a sentence how it was taken. Khan attributes the release to an unknown actor, presumably the one who exploited the same issue, and writes that a different person found his proof of concept on his test repository and used it against Cline directly. Cline’s post-mortem thanks him for identifying indicators of compromise in its nightly workflows.
step by step
- 21 December 2025an AI bot starts sorting issuesCline adds a GitHub Actions workflow in which Claude reads each new issue, with shell access and read-only rights to the code.
- 1 January to 9 February 2026a researcher reports, then publishesAdnan Khan reports that any issue title can make the bot run code and reach the release pipeline, follows up three times, and publishes on 9 February.
- 9 February 2026the fix and the missed tokenCline removes the workflow within about half an hour and rotates its tokens, but deletes the wrong npm token.
- 17 February 2026an unauthorised release goes outAn unknown party uses the still-active token to publish cline@2.3.0, which installs another AI agent on every machine that runs its install script.the assumption breaks here
- 24 February 2026the post-mortemCline removes its AI triage workflows, moves npm publishing to short-lived credentials and explains the rotation failure.
how it surfaced
StepSecurity, a vendor that sells supply-chain monitoring, wrote on 17 February 2026 that its systems flagged the release within about a quarter of an hour, by its own timeline. The signals it lists: the version was published by hand from an npm user account, outside the project’s usual publishing from GitHub Actions, and without a provenance attestation. StepSecurity also wrote that the version was downloaded approximately 4,000 times, without saying how it counted. A download count does not confirm installs, and no source gives a number of affected machines.
Cline published version 2.4.0, deprecated 2.3.0 and revoked the token at 11:23 Pacific time, about eight hours after the release. The advisory tells users to update and remove OpenClaw if it was installed unintentionally. Cline describes the payload as benign and the impact on users as low. That is the company’s assessment of a package that installed, without the developer asking, an AI agent with access to whatever the developer’s machine could reach.
The post-mortem lists the changes. AI triage workflows are gone, and any future automation there is to be read-only with no shell. Publish workflows no longer use the cache. npm publishing now uses short-lived credentials issued to GitHub Actions, with provenance, so no long-lived npm token exists to leak. Rotations are to be checked on the credential itself, by confirming the old one no longer works.
The Nx packages of August 2025 were reached through a similar opening, a pull request title that ended up inside a shell command, with no AI in between. The Amazon Q extension case is a build token that wrote straight into a release. The guide on agent permissions covers why a prompt widens what an agent can do, pipeline secrets covers the publish tokens a build holds and how short-lived ones replace them, and the coding agent guide covers agents that run with whatever their machine holds.
An AI agent that reads text from strangers can do whatever the machine it runs on can reach.
Pick one bot or assistant in your company that reads messages from outside and list what it could run if a message told it to.
picture it at a company like yours
The company here is made up. The other two paths are not: they happened, and each links to the story.
- in this storythe day it mattersA stranger sends a request whose subject line tells the assistant to run something, and it does.what it costs youThe keys behind your invoices work for someone else now, and you find out when a customer asks about a strange email.
- or it goes like thisit happenedEchoLeak / Microsoft 365 Copilot, 2025howAn email with hidden instructions reaches an assistant that reads every inbox.what leavesanything the assistant can seewho ends up with itthe stranger who sent the email
- or it goes like thisit happenedMCP tool poisoning, 2025howA tool description tells the agent to send files somewhere else.what leavesfiles the agent can openwho ends up with itwhoever wrote the tool
sources
- Cline: post-mortem, unauthorized Cline CLI npm publish, 24 February 2026
- GitHub advisory GHSA-9ppg-jx86-fqw7: Unauthorized npm publish of Cline CLI cline@2.3.0 with modified postinstall script to install openclaw, cline/cline, 17 February 2026
- Adnan Khan: Clinejection, 9 February 2026, updated to 18 February 2026
- StepSecurity: Cline supply chain attack detected, cline 2.3.0 silently installs OpenClaw, 17 February 2026
- Anthropic: claude-code-action security documentation (read 13 November 2026)
your turn
what you can do about it
this week
List every AI bot that reads outside messages, and remove its shell and any key it does not need.
Go through your support inbox, ticketing tool, chat channels and code repositories, and note each AI assistant or bot that reads text from customers or the public. For each one, write down which tools it may use (running commands, writing files, fetching pages) and which keys sit on the machine or account it runs under. Cline's bot only needed to read and label issues, and it had a shell on a machine that shared a cache with the release: which of yours can do more than its job needs?
go deeper
if you're not sure where to start
Run the offboarding checklist on one leaver. It covers every account and token to close when someone leaves.
open the checklist →
