← Voltar ao feed
medium technology

Claude Code VS Code extension permission prompts hide symlink targets, enabling workspace escape; Anthropic closed report as informative

1 fonte(s) · 1 atualização(ões) · publicado 16/11/2026 09:58 · coletado há 1 dia

A researcher (Ultrastrike) reports that Anthropic's Claude Code extension for Visual Studio Code shows only the workspace-relative path in permission prompts. When a path goes through a symlink that points outside the workspace, the prompt does not show the real destination. A rogue or prompt-injected agent could therefore get approval for reads or writes outside the trusted workspace. The researcher says the issue is unpatched in plugin version 2.1.322. Anthropic closed the report as informative and did not treat it as a vulnerability. No CVE is cited, and this analysis did not include web research, so vendor severity and advisory details are unverified.

Entenda a falha

A vulnerabilidade

No CVE has been assigned. The Claude Code extension for Visual Studio Code, version 2.1.322, shows only the workspace-relative path in its permission prompts. If that path goes through a symlink pointing outside the workspace, the prompt does not show the real destination. A rogue or prompt-injected agent could get a user to approve reads or writes outside the trusted workspace, for example into ~/.ssh/ if a symlink points at a home directory. The researcher says it is unpatched in 2.1.322, and no fixed extension version is cited. Anthropic closed the report as informative and did not treat it as a vulnerability.

Como funciona

The agent touches a file through a symlink inside the workspace, such as ./helpdocs, which points elsewhere. The VS Code plugin asks the user for approval, but the prompt shows only the in-workspace path. In testing, a write of test.txt through that link gave no sign that the file would land outside the project. The user approves based on inaccurate context, and the operation then acts on the real target outside the workspace. A symlink can reach the workspace through a malicious Git commit, or an agent could rename a familiar symlink, since the plugin may rename workspace files without approval depending on the permission mode.

Por que é vulnerável

The prompt does not resolve symlinks before asking for a permission decision and does not show the canonical destination, so the displayed path misrepresents where the operation lands. According to the researcher, Anthropic argued that the prompt's displayed details are not part of the security control and that symlinks fall outside Claude Code's permission threat model. The researcher says the public HackerOne guide does not list that exclusion and calls it a UI/UX security issue. The Claude Code CLI was later hardened to resolve symlinks and show the real destination, but that fix never reached the extension.

Como foi descoberta

A researcher known as Ultrastrike reported the issue to Anthropic, which closed it as informative; the researcher disagrees. The finding builds on Adversa AI's May 2026 'SymJack' research, which Adversa says worked against Claude Code, Gemini CLI/Antigravity CLI, Cursor Agent CLI, GitHub Copilot CLI, Grok Build and OpenAI Codex CLI. Adversa says Anthropic rejected its own report as out of scope, then hardened the CLI a few weeks later. Ultrastrike found that this fix never reached the extension. The articles do not say when the extension report was filed.

PoC existente, privada

O exploit existe, mas quem o tem não o publicou.

The researcher showed in testing that a write of test.txt through a symlink gave the user no sign that the file would land outside the project. This is a verification of the misleading prompt only. The articles cite no published exploit or repository, so the demonstration appears to be kept by the researcher. Because the exploit depends on a symlink existing and on the user approving the prompt, urgency is moderate, and checking paths manually with realpath is the available mitigation.

Explicação gerada por IA a partir da matéria e dos artigos, em 16/11/2026 10:16.

Detalhes técnicos

  • A symlink such as ./helpdocs can point elsewhere but appears as an in-workspace path. In testing, a write of test.txt through it gave no sign that the file would land outside the project.
  • Impact depends on which symlinks exist. A link to a home directory could expose ~/.ssh/ keys.
  • The researcher says the finding builds on Adversa AI's May 2026 "SymJack" research. Adversa reports that the technique works against Claude Code, Gemini CLI/Antigravity CLI, Cursor Agent CLI, GitHub Copilot CLI, Grok Build and OpenAI Codex CLI. In that research, a malicious MCP server then runs attacker code on the next restart.
  • Two scenarios weaken the trusted-workspace assumption. Git supports symlinks, so a malicious commit from a compromised account could deliver one to colleagues. Depending on the permission mode, the plugin may also rename workspace files without approval, so an agent could rename a familiar symlink.
  • Anthropic reportedly argued that the prompt's displayed details are not part of the security control and that symlinks are outside its permission threat model. The researcher says the public HackerOne guide does not list that exclusion.
  • The Claude Code CLI was later hardened to resolve symlinks and show the real destination. According to the researcher, that fix never reached the extension, which has over 30 million Marketplace installs.

Remediação

  • No fix exists for the extension. Before approving a request, run realpath on the displayed path. Deny the request if it resolves outside the workspace or other allowed locations.
  • Adversa recommends that vendors resolve symlinks before any permission decision and show the canonical destination in the prompt.
  • Organizations should scan repositories for tracked symlinks that target configuration files and fail pull requests that modify agent or MCP configuration access. They should also isolate CI runners that handle untrusted code from those holding production secrets.

Linha do tempo

  1. gbhackers.com ↗ · pub. 16/11/2026 09:58 · coletado 16/11 10:00 primeiro registro

    Claude Code VS Code extension permission prompts hide symlink targets, enabling workspace escape; Anthropic closed report as informative

    ler a análise desta fonte

    A researcher (Ultrastrike) reports that Anthropic's Claude Code extension for Visual Studio Code shows only the workspace-relative path in permission prompts. When a path goes through a symlink that points outside the workspace, the prompt does not show the real destination. A rogue or prompt-injected agent could therefore get approval for reads or writes outside the trusted workspace. The researcher says the issue is unpatched in plugin version 2.1.322. Anthropic closed the report as informative and did not treat it as a vulnerability. No CVE is cited, and this analysis did not include web research, so vendor severity and advisory details are unverified.

    Technical Details

    • A symlink such as ./helpdocs can point elsewhere but appears as an in-workspace path. In testing, a write of test.txt through it gave no sign that the file would land outside the project.
    • Impact depends on which symlinks exist. A link to a home directory could expose ~/.ssh/ keys.
    • The researcher says the finding builds on Adversa AI's May 2026 "SymJack" research. Adversa reports that the technique works against Claude Code, Gemini CLI/Antigravity CLI, Cursor Agent CLI, GitHub Copilot CLI, Grok Build and OpenAI Codex CLI. In that research, a malicious MCP server then runs attacker code on the next restart.
    • Two scenarios weaken the trusted-workspace assumption. Git supports symlinks, so a malicious commit from a compromised account could deliver one to colleagues. Depending on the permission mode, the plugin may also rename workspace files without approval, so an agent could rename a familiar symlink.
    • Anthropic reportedly argued that the prompt's displayed details are not part of the security control and that symlinks are outside its permission threat model. The researcher says the public HackerOne guide does not list that exclusion.
    • The Claude Code CLI was later hardened to resolve symlinks and show the real destination. According to the researcher, that fix never reached the extension, which has over 30 million Marketplace installs.

    Remediation

    • No fix exists for the extension. Before approving a request, run realpath on the displayed path. Deny the request if it resolves outside the workspace or other allowed locations.
    • Adversa recommends that vendors resolve symlinks before any permission decision and show the canonical destination in the prompt.
    • Organizations should scan repositories for tracked symlinks that target configuration files and fail pull requests that modify agent or MCP configuration access. They should also isolate CI runners that handle untrusted code from those holding production secrets.