Your AI coding agent will read anything you drop in Downloads

October 26, 2026 · SPR{K3 Research

A security researcher drops a text file into your Downloads folder. The next time you ask your AI coding agent to help with something and it starts up in that folder, it reads the file and runs the commands inside. No prompt. No click. No warning.

That is the attack Forever Security's Gal Weizman disclosed to BleepingComputer on October 25, and it works against the three most-used AI coding agents on developer machines today: Claude Code, OpenAI Codex, and Cursor CLI.

The pattern

All three agents look for a conventional instructions file in their working directory on startup and treat it as authoritative local configuration. The file is called AGENTS.md, or CLAUDE.md when the agent is Anthropic's. The convention is public and documented — Nvidia's own developer blog called out the indirect-prompt-injection risk earlier this year.

Attackers do not need to reach your source tree, your repo, or your package manager. They need to reach one folder on your laptop. The folder every major browser writes to by default is ~/Downloads, and Weizman told BleepingComputer that hundreds of Forever's customers run coding agents from there — Claude Code alone launched over 1,300 sessions from that path across Forever's observed base.

What Forever observed and tested

Forever Security caught this in a real customer environment. An AGENTS.md file ended up in a developer's Downloads folder; when the agent started there, it tried to install a malicious npm package named function-color. Forever could not determine how the file arrived on the host.

In controlled testing, Forever got the following numbers:

Chrome, Edge, and Firefox save downloads to ~/Downloads by default and show only a notification. Brave, Weizman notes, prompts per-download.

No one is going to patch this for you

Here is the hard part. Weizman is explicit in the article: "There is nothing here for OpenAI or Anthropic to patch." The behavior is intended browser semantics meeting intended agent semantics. No CVE has been assigned. No vendor advisory is on the way.

That matters for the week it lands in. In the last five days we have watched four different AI-infrastructure vendors tell their own users that in-the-wild attack surfaces are not the vendor's responsibility: Anomaly called malicious MCP servers "outside our trust boundary" when Pillar Security disclosed a command-injection in OpenCode's OAuth flow; OpenAI classed a working cross-repo-write prompt-injection in the ChatGPT GitHub Connector as "out of scope" via The Hacker News ThreatsDay; the x402 Foundation shipped silent fixes for three critical flaws in its agent payment SDK without issuing CVEs or advisories; and now Forever's finding adds AGENTS.md hijack to the list.

The agent-config boundary is where the enterprise defender is on their own.

What the agent is actually doing

The agent reads a file on your filesystem that it treats as authoritative configuration, and converts the file's contents into tool calls — shell commands, package installs, API traffic rewrites — without a human approval gate at the per-command boundary. That is the runtime pattern. Whether the file arrived via a drive-by download, a cloned repo, a shared folder, a scp from a compromised peer, or anything else is secondary.

Two of the three outcomes Forever measured land in cleanly familiar categories. The npm install of function-color is a straight supply-chain compromise: attacker-authored instructions become a dependency fetch, exactly the pattern we saw with the LiteLLM 1.107.0 PyPI compromise shipping the "privcollect" stealer last week and the MALFEX npm chain disclosed by SafeDep in mid-October. The CLAUDE.md API-redirect is a man-in-the-agent primitive: once the attacker's server impersonates the model, it can hand the agent arbitrary tool calls on arbitrary code, and the agent will run them because they look like model output.

What to do today

Forever's guidance, which lines up with what you would do on any other developer-workstation hygiene problem, is specific:

None of this is new advice. All of it is advice developers routinely skip, because the agent's job is to make moving fast feel safe. The point of this disclosure is that the agent cannot tell the difference between a legitimate AGENTS.md you wrote and a hostile one someone else put next to it. You are the only one who can.


SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.