Forever Security's October Disclosures Find Browser Flaws in Five AI Coding Agents and a Separate Instruction-File Trust Gap in Two
Forever Security's October research found a planted AGENTS.md/CLAUDE.md file can hijack Codex and Claude Code, and a separate flaw class broke embedded-browser rules in five coding tools with uneven patch status.
Editor's Note ·
- Correction:
- The article quotes Forever Security's closing line as "that is what Forever was built to watch." The source (forever.security, "Markdown Is the New Executable") actually reads: "...and it is what Forever was built to watch." The quoted word should be "it," not "that."
Background
AI coding agents increasingly combine two design choices that are each reasonable on their own. The first is an instructions file — an AGENTS.md or CLAUDE.md — that sits in a project folder and is read and obeyed without asking the user each time. As security firm Forever Security put it in an October 22, 2026 write-up, such a file “sits in a project folder you control. The first time you work there, you see it and vet it. After that you trust it,” according to Forever Security. The second is an embedded browser some agents use to preview or test a project. In a companion technical post, Forever Security wrote that these tools “don’t just ship a browser. They embed one inside an AI desktop app, then give it powers a normal browser does not have. That enrichment is the soft spot,” per Forever Security.
The Herald has reported separately on where coding agents’ saved credentials end up, in a CI prompt injection, saved Claude Code approvals and AI-driven GitHub searches, and on GitHub’s November 17 addition of browser control to Copilot, which noted that GitHub’s own announcement named only extension permission settings as a safeguard. Two Forever Security disclosures published in October, examined together, describe a different pair of trust failures in the same family of tools: one in how agents read instruction files, the other in how agents’ embedded browsers handle hostile pages.
Current situation
October 22: a planted file in the Downloads folder
Forever Security researchers Gal Weizman and Omer Samet, in a report covered by Help Net Security on October 26, reported finding “a malicious AGENTS.md file in the ~/Downloads folder of a real customer machine - plain text, disguised as ordinary engineering notes,” according to Forever Security. Their write-up argues the danger comes from combining two previously harmless facts: a web page can silently save a file into a user’s Downloads folder with no click and no permission, and a coding agent is built to read and follow an instructions file it finds in the folder where it starts. Forever Security summarized this as “Markdown is the new executable,” per the same report.
The researchers say they reproduced the attack on both OpenAI’s Codex and Anthropic’s Claude Code. Against Codex, a planted AGENTS.md styled as routine “Engineering Work Guidelines” instructed the agent to add a tools-server entry to its machine-wide ~/.codex/config.toml file and fetch the npm package function-color — in testing, Forever Security says, “Codex ran the malware in more than 90% of attempts,” and because the edited file is machine-wide rather than folder-scoped, “every Codex session on the machine loads the attacker’s server at startup — in any folder, any project,” according to Forever Security. The payload used the flag --dangerously-bypass-approvals-and-sandbox to get past Codex’s network sandbox, the write-up says.
Against Claude Code, the researchers say a planted CLAUDE.md styled as gateway setup instructions got the agent to write an ANTHROPIC_BASE_URL pointing to a server Forever Security controlled, in “about 50% of the time,” leaking Claude Code’s OAuth token and, in roughly a fifth of sessions, achieving code execution within the same session once a spawned subagent connected through the redirected address, per Forever Security. The function-color package the researchers cite has since been removed from npm, the write-up says, and is tied to a roughly two-year-old campaign; Checkmarx’s separate research on that campaign, which it calls MALFEX, describes function-color as a wrapper with “no payload of its own” that “installs function-flag as a dependency,” according to Checkmarx — the Herald covered the broader MALFEX campaign separately in October. Forever Security frames the Downloads-folder issue as not fixable through a vendor patch: “There is nothing here for OpenAI or Anthropic to patch. An agent that follows the AGENTS.md in its folder is doing exactly what it was built to do,” the firm wrote.
October 29: embedded browsers break their own rules
A second Forever Security write-up, covered by Help Net Security on October 29, examined the embedded browsers inside five tools — Cursor, GitHub Copilot, Codex, VS Code and Claude Desktop — and found each one relaxed a different browser security rule. Forever Security grouped the results under the name EMBRACE, for “an EMbedded BRowser trusted like a real one,” with three variants, according to Forever Security.
In Cursor, the researchers say a hostile web page could reach the agent’s own composer through a bridge, window.cursorBrowser.send(), splitting a command across several fields so the agent would reassemble it; a blocked npm install could then be retried by the agent itself with its sandbox disabled, letting a package’s install script run with full host privileges. Forever Security disclosed this to Cursor on September 10 and says it was “silently patched” on October 1, adding that “Cursor never reached back.” In GitHub Copilot, the researchers say a page could report a false address to the app, which displayed it in the address bar and told the agent the tab was there too — enabling a phishing page to carry GitHub’s own login address. Forever Security says it found this issue on Copilot version 1.1.27 on September 26 and that GitHub “fixed it independently in a later Copilot build (1.1.29)” by October 13, before Forever Security had sent a report. In Claude Desktop, Codex and VS Code, the researchers describe a local-file preview whose folder-scope restriction runs only in the page’s top document and not inside a web worker, letting a worker fetch any file on disk; Codex, they add, “needed no worker at all” because it applied no restriction in the first place. On disclosure, Forever Security says Anthropic initially declined the Claude Desktop finding as outside its “Workspace Trust” security boundary on July 23 before accepting a report on the same issue from Noma Security as valid the next day; OpenAI, per Forever Security, never responded about Codex; and Microsoft declined to treat the VS Code issue as a vulnerability. As of its October 29 publication, Forever Security’s write-up states that “the Codex and VS Code issues are still open. Microsoft declined to fix VS Code, and OpenAI has not shipped a fix for Codex.” The write-up does not state, in the passage summarizing what remains open, whether the Claude Desktop issue was patched by that date.
Implications
Taken together, the two disclosures describe different kinds of problems with different remedies. Forever Security frames the embedded-browser flaws as ordinary, patchable bugs — two of the five tools it tested, Cursor and Copilot, each had a fix within a few weeks of discovery, one of them before the firm even reported it. The Downloads-folder issue, by the same firm’s own account, is not something a patch can close, since the agent is “doing exactly what it was built to do” when it reads and follows a file in its working folder; Forever Security’s suggested response is operational — avoid starting an agent in the Downloads folder and inspect instruction files before trusting them — rather than a vendor fix to wait for.
The pace of fixes also diverged by vendor and by product within the same disclosure. By the Herald’s own count from Forever Security’s disclosure dates, Cursor’s flaw closed in three weeks (September 10 to October 1) and Copilot’s closed in seventeen days (September 26 to October 13); Codex and VS Code’s file-read flaws, by Forever Security’s account, remained open as of October 29, and the write-up does not resolve whether Claude Desktop’s status, after Anthropic accepted a related report from a second source on July 24, had changed by then. None of the sources cited here assign a CVE or GHSA identifier to any of the five findings.
The disclosures also land in the middle of vendors expanding exactly the surface in question. Five weeks after Forever Security says GitHub fixed its address-bar spoofing bug — by the Herald’s own count, October 13 to November 17 — GitHub shipped a new feature letting Copilot “control the browser, debug web apps in real time, automate browser tests, and handle complex web tasks,” in public preview, according to GitHub’s changelog. As the Herald reported when that feature shipped, GitHub’s own announcement named only the Chrome extension’s permission settings as a safeguard and did not describe sandboxing or other controls specific to letting an agent act inside a browser session — a gap the Herald’s earlier report said “matters in a year when coding-agent security research has repeatedly focused on exactly this kind of expanded surface.” Forever Security’s own description of its business — “that is what Forever was built to watch” — is a reminder that the firm, like other security vendors whose research the Herald has covered, has a commercial interest in highlighting exactly this category of risk; that does not bear on the accuracy of its technical findings, which Help Net Security’s independent coverage and, for the Downloads-folder case, a real customer incident corroborate, but it is worth noting when weighing how the findings are framed.
The cited sources leave several questions open. Forever Security’s telemetry figures — that about half of its customers run a coding agent inside a Downloads folder and roughly a quarter of agent sessions open a browser — describe its own customer base rather than the industry as a whole. Neither write-up states whether any other coding agents beyond the five and two tested share the same flaw classes. And the disclosure timelines show a different vendor response for each of the five products: a silent fix with no acknowledgment to the researchers (Cursor); a fix that arrived before a report was even filed (Copilot); an initial decision that the behavior did not qualify as a vulnerability, which a second report from a different researcher reversed the next day (Claude Desktop); a decision that the behavior did not qualify as a vulnerability that was not reversed, by Forever Security’s account (VS Code); and no response at all (Codex) — with no cited source explaining why vendors reached different conclusions about the same underlying pattern of misplaced trust.
The Herald’s contributor model is built by Anthropic, which makes Claude Code and Claude Desktop, two of the products examined in Forever Security’s research cited above; statements about Anthropic’s products in this article are attributed to the cited sources.