Ultrastrike researcher Carl Pearson reported on November 11, 2026 that Anthropic's official Claude Code VS Code extension can show permission dialogs that hide where symlinks really point. If you approve a path that looks inside the project, the agent may read or write outside the trusted workspace. Anthropic closed the HackerOne report as informative and does not treat it as a vulnerability, per Ultrastrike. No patch was available in extension 2.1.322 at publication. Until the UI matches the CLI symlink warnings, run realpath on every non-obvious approval path and deny requests that resolve outside your workspace.
| Question | Answer |
|---|---|
| Am I affected? | You are if you use the Claude Code VS Code extension (29M+ installs per Ultrastrike) and your tree contains symlinks, including ones synced from Git. |
| Is there a fix? | Not in the extension at 2.1.322 per Ultrastrike; CLI behavior was updated after Adversa AI's May 2026 research. |
| Is this the Edit TOCTOU CVE? | No. CVE-2026-88968 is a separate macOS Edit race patched in 2.1.251. |
| What should I do before approving? | Copy the dialog path and run realpath <path>; deny if the result leaves the workspace. |
| Does data leave my machine? | This is a local filesystem boundary question: approvals may touch ~/.ssh, dotfiles, or other host paths reachable through symlinks. |
Adversa’s May 2026 baseline: five agents, one UX class
Before Ultrastrike’s VS Code retest, Adversa AI documented the same display mismatch across Claude Code CLI, Cursor, Copilot, Grok Build, and related harnesses. The pattern is consistent: the product detects that a symlink touches sensitive territory, asks for human approval, but prints a workspace-relative alias instead of the resolved path.
Adversa’s ./link → /etc/passwd example is the mental model. A developer sees ./link/passwd and reasonably assumes the write stays in the repo. The kernel follows the symlink to /etc/passwd. That is not remote code execution by itself, but it breaks the story the UI tells at the moment of consent.
Anthropic initially treated Adversa’s CLI report as non-security, per both Adversa and Ultrastrike. Adversa says behavior changed quietly on CLI weeks later. Ultrastrike’s November work asks a narrower question: did the VS Code shim inherit that CLI fix? The answer, for 2.1.322, was no.
Security teams should file this under CWE-451-style misleading UI and CWE-59 symlink handling, adjacent to—but not identical to—CVE-2026-88968, which is a race inside Edit validation rather than a dialog labeling bug.
What Ultrastrike found
Pearson tested the marketplace extension while reviewing VS Code plugins. He expected parity with CLI fixes after Adversa's disclosure across Claude Code, Cursor, Copilot, Grok, and related tools.
In his reproduction, ./helpdocs inside the workspace was a symlink to a location outside the project. When the agent asked to write test.txt under that path, the permission UI showed only the workspace-relative route. Nothing in the dialog indicated the file would land outside the trusted folder.
Image: Ultrastrike Cybersecurity
Ultrastrike's post includes a second screenshot comparing Claude Code CLI: the terminal build shows the resolved path and warns when a symlink resolves outside a trusted folder. That matches what Cyber Security News summarized on November 16, 2026 after Adversa's earlier ./link → /etc/passwd style example, where the prompt displayed ./link/passwd instead of the real target.
Attack chain step by step
This is a consent manipulation chain, not remote code execution by itself.
- Symlink present. A folder or file in the workspace is a symlink. It may arrive through Git (Git stores symlinks), a dependency, or an attacker-controlled commit after an account compromise, as Pearson describes.
- Agent requests access. Claude Code in VS Code asks the developer to approve a read or write on a path that appears under the project root, for example
helpdocs/test.txt. - Dialog omits resolution. The extension shows the workspace-relative string. Ultrastrike says it does not show the canonical target or an outside-workspace warning at 2.1.322.
- User approves. A developer who trusts the visible path clicks allow.
- Operation hits the real path. The tool follows the symlink. Ultrastrike notes impact depends on the link target, with
~/.sshcalled out if a home-directory symlink exists.
Prompt injection or a misaligned agent decision fits the same chain: the model does not need a kernel bug if the human approves a misleading path. That pattern sits beside other permission UX failures Beam has covered, from embedded browser bridges in IDEs to the compromised Claude Code 2.1.313 extension that abused marketplace trust entirely differently.
What Anthropic said
Pearson filed through HackerOne on October 22, 2026, per Ultrastrike. Anthropic closed the report as informative, arguing the permission prompt itself is sufficient and that text in the dialog is not part of the security control. Anthropic also said symlinks are out of scope for the permission threat model, although Ultrastrike notes the public program guide does not clearly list that exclusion.
Ultrastrike disagrees: if opening a workspace fully trusts symlinks, prompting on symlink traversal implies a boundary that only works when the user sees accurate destinations.
Does Claude Code show symlink targets in the CLI today?
Ultrastrike’s comparison screenshot is the reference point for expected behavior after the Adversa cycle. When the terminal build requests access through an outside symlink, the user should see:
- The canonical path the operation will touch.
- An explicit warning that resolution leaves a trusted folder.
That is the bar Pearson wants in VS Code. Until the extension matches it, treat IDE and CLI as different trust surfaces even when both say “Claude Code.” A team might pin CLI 2.1.251+ for TOCTOU while an engineer still runs a VS Code extension that omits symlink context.
Extension version checks are separate from claude --version. In VS Code, open the Extensions panel, select Anthropic Claude Code, and read the installed version against Ultrastrike’s 2.1.322 finding. After any update, repeat the realpath habit on the first symlink-heavy repo you open.
Repository symlink hygiene for agent users
Git stores symlinks as tree objects. They survive git clone and git pull exactly like files, which makes them a supply-chain persistence mechanism inside an otherwise normal commit.
Practical checks for teams that let agents operate on shared monorepos:
git ls-files -s | awk '$1 ~ /^12/ {print $4}'
List symlinks tracked in the index, then inspect targets:
readlink -f path/from/index
Require CODEOWNERS review for new symlinks in agent-enabled repos, the same way you would for install scripts or .github/workflows changes. Pair that with MCP config review so a poisoned skill does not ask the agent to create the symlink in the first place.
If your org uses Cursor or forks that repackage the Anthropic extension, map findings back to the underlying Claude Code build after incidents such as the tampered 2.1.313 marketplace cache—a separate problem, but part of the same IDE trust surface.
Trusted workspace limits Pearson highlights
Two scenarios extend beyond a one-off malicious repo:
Shared Git history. A compromised teammate account can push a commit with a symlink. After git pull, every engineer's previously trusted tree includes the link.
Renamed symlinks. Depending on permission mode, the extension may rename files without asking. A known symlink name can disappear from the next dialog (./foo/bar → ./foo/bat), so the user no longer recognizes the link they learned to reject.
Those are reasons to treat approval dialogs as security decisions, not annoyances, the same way we treat MCP config changes before enabling new tools.
What this means if you run AI coding agents
Patch what has CVEs first. Upgrade Claude Code CLI past 2.1.251 for CVE-2026-88968 with claude --version and your package manager pin. The VS Code display issue is orthogonal to that Edit TOCTOU fix.
Use the Ultrastrike workaround on every suspicious approval:
realpath <path shown in permission dialog>
If the resolved path is outside the workspace (and outside any path you intentionally allow), deny and inspect intermediates for symlinks.
Audit symlinks after sync:
find . -type l -ls
Tighten permission modes so agents cannot rename paths you rely on to recognize links. Review Anthropic's VS Code permission settings after each extension update.
Separate untrusted trees. Disposable clones for unknown OSS limit symlink blast radius, consistent with guidance in CVE volume and agent patch discipline.
Document a deny-by-default habit for symlink approvals. Security questionnaires increasingly ask whether AI tools can touch dotfiles, SSH keys, or cloud credential paths. “We read the dialog” is not a control if the dialog omits resolution. Record realpath checks in your internal runbook the way you record two-person review for production deploy keys.
Revisit after Anthropic ships UI parity. When release notes mention symlink display, re-run Ultrastrike’s ./helpdocs style test in a throwaway workspace before you relax manual checks org-wide.
How this differs from embedded-browser and MCP bridges
Pearson’s issue stays inside the filesystem permission model. It is not the same as embedded browser flaws where web content reaches agent context, or MCP rug pulls where a server swaps tool definitions after install.
The common thread is consent fatigue: developers click through prompts to stay productive. Any layer that hides the object being touched—whether a URL, an MCP tool schema, or a symlink target—turns approvals into rubber stamps. Local observability helps after the fact; accurate prompts help at the moment of decision.
What Beam can and cannot do here
Beam does not fix Anthropic's dialog text and does not intercept VS Code's native permission UI.
On the same machine, Beam still helps coding-agent operators:
- PreToolUse hooks on Claude Code and Codex can deny matched high-risk tool patterns when workspace policy is in enforce mode — documented in agent hooks. That can block some exfiltration or destructive commands after a bad approval, but not every read/write the IDE allows through its own prompt.
- Heuristic scans of skills and MCP configs (11 patterns plus MCP version-pin checks) flag risky install shapes before sessions start — see scanning.
- Local event capture with redaction before storage supports post-incident review: what did the agent propose around the time a symlink path was approved?
Missing or expired cached policy is fail-open on hooks. Beam is not a kernel sandbox and does not re-resolve every path the VS Code extension displays.
Related reading
- Claude Code CVE-2026-88968 Edit TOCTOU on macOS
- Compromised Claude Code VS Code 2.1.313 extension
- MCP security practical guide
- Stopping dangerous Claude Code commands
- What is AI agent monitoring
- Beam: local-first agent security
Frequently asked questions
Does the Claude Code VS Code extension show real symlink targets?
Ultrastrike reports that as of extension version 2.1.322, the permission dialog still shows workspace-relative paths without the resolved location or an outside-workspace warning, unlike the current Claude Code CLI after Adversa AI's May 2026 disclosure.
Is the Claude Code VS Code symlink issue a CVE?
There is no published CVE for this behavior. Anthropic closed Ultrastrike's HackerOne report as informative and stated symlinks are outside the permission threat model, per Ultrastrike's November 11, 2026 write-up.
How do I check a Claude Code VS Code approval path?
Before approving, run realpath on the path shown in the dialog. If the resolved location is outside your workspace or other allowed roots, deny the request and inspect the path for symlinks, per Ultrastrike's workaround.
Did Anthropic fix this in Claude Code CLI?
Adversa AI reported misleading symlink prompts in CLI tools in May 2026. Ultrastrike and Cyber Security News say the CLI now displays the actual file location and warns when a symlink resolves outside a trusted folder; the VS Code extension had not received equivalent UI at 2.1.322.
Can prompt injection use symlink approvals?
Ultrastrike states a rogue or prompt-influenced agent could ask for access through a symlink that looks safe in the dialog, with impact depending on what the link targets, such as SSH keys under ~/.ssh if a home symlink exists.
How is this different from CVE-2026-88968?
CVE-2026-88968 is a TOCTOU race in Claude Code's Edit tool on macOS before 2.1.251. The VS Code issue is a display gap in permission prompts, not the same patched Edit validation bug.
Is Claude Code safe in VS Code?
Safety depends on harness version, permission mode, and what is in the workspace. Ultrastrike's finding means you should not trust the dialog path alone when symlinks exist; verify with realpath until Anthropic ships CLI-parity warnings in the extension.
Version numbers, extension behavior, and vendor responses are accurate as of November 16, 2026 per Ultrastrike, Adversa AI, and Cyber Security News primary reporting.

