Security researchers have found that the Claude Code extension for Visual Studio Code can show a permission prompt that hides where a file really sits.
A user may approve access to a path that looks safe, while the file lives outside the trusted workspace. Anthropic reviewed the report and does not treat it as a security issue.
The extension has more than 29 million installs, so the problem could reach a very large number of developers. The extension is Anthropic’s official add-on that lets developers use Claude Code inside the VS Code editor. It asks for approval before the AI agent reads or writes files, and that prompt is meant to protect the user.
The same type of weakness was seen before in the command-line version of the tool. A Claude Code symlink import issue reported earlier also showed how a link inside a project can quietly point to files outside it.
The new finding shows the editor extension still has a similar gap. Researchers at Ultrastrike identified and described the issue in a blog post published on November 11, 2026. Carl Pearson wrote the report. He found it while reviewing several VS Code plugins for flaws.
Claude Code VS Code Extension Can Access Files
The problem centers on symlinks, which are shortcuts that point to another file or folder. In the test, a folder named ./helpdocs inside the workspace was a symlink to a location outside it.
When the agent asked to write a file called test.txt, the prompt showed only the workspace path. It gave no sign that the file would land elsewhere.
The roots of this go back to May 2026. Adversa AI published research showing that the Claude Code command-line tool had the same display flaw.
A link at ./link that pointed to /etc/ would show as ./link/passwd in the prompt, not the real /etc/passwd. Anthropic reportedly called that report not a security issue, but Adversa says the company fixed it quietly a few weeks later.
The command-line tool now shows the real file location. It also warns the user when a symlink leads outside a trusted folder. The VS Code extension never got the same change. Pearson says it still fails to show the link target in version 2.1.322, the latest at the time of writing.
The risk grows when an agent is tricked. A rogue agent, or one fooled by a malicious instruction, could get a user to approve access that seems harmless.
Attackers already target tools like this, as seen in how attackers can hijack AI coding agents with planted text. The real impact depends on which symlinks exist. A link to the home folder, for example, could expose SSH keys stored in ~/.ssh/.
Anthropic’s Trusted Workspace Argument
Pearson reported the issue to Anthropic through its HackerOne program. Anthropic closed the report as informative. It said the prompt alone is enough and that the text shown in it is not part of the security control.
It also said symlinks fall outside the permission threat model, although Ultrastrike says the public program guide does not list that exclusion.
Anthropic’s view rests on a trust model. Once a user opens a workspace, the tool trusts everything in it, including symlinks. Pearson questions this logic.
If the workspace is fully trusted, he asks, why does the tool ask for permission at all? A prompt suggests a security boundary, and that boundary only works if the user sees accurate details.
He also points to two cases where the trust model can fail. First, a developer’s account in the same company may be compromised. The attacker can push a bad commit with a symlink, and it lands in every teammate’s workspace on the next sync, because Git supports symlinks.
Second, depending on the permission mode, the extension can rename files without asking. A known link could be renamed, and the user would no longer spot it in the prompt.
These risks matter more as teams lean on AI tools. Other recent research on AI coding agent flaws shows that weak guardrails around permissions are a repeated pattern, not a one-off bug.
No patch exists yet, so Ultrastrike suggests a manual check. Before approving a request, copy the path from the prompt and run realpath in a terminal.
The command prints the true location on disk. If that location is outside the workspace, the agent’s request may involve a symlink and should be denied.
Developers should also review repositories for unexpected symlinks after each sync. Teams should be careful with extensions too, since a malicious Claude Code extension recently appeared in extension marketplaces and showed how much trust these tools carry.
Indicators of compromise (IoCs):-
| Indicator | Type | Detail |
|---|---|---|
| Claude Code VS Code extension 2.1.322 | Affected version | Latest version, unpatched at publication |
./helpdocs | Example symlink | Points outside the workspace |
./link pointing to /etc/ | Example symlink | Prompt shows ./link/passwd, not /etc/passwd |
~/.ssh/ | Exposed path | SSH keys reachable through a home folder symlink |
realpath | Check command | Reveals the real file location |
Note: IP addresses and domains are intentionally Qdefanged (e.g., [.]) to prevent accidental resolution or hyperlinking. Re-fang only within controlled threat intelligence platforms such as MISP, VirusTotal, or your SIEM.
Give your SOC the alert context on the latest threats to stop incidents quickly -> Integrate Threat Intel from 16K SOCs

