SNEAKPEEK — CVSS 8.7 — PATCHED AUG 17

Is your Claude Code build patched against SneakPeek?

Security researchers found that Claude Code Desktop's preview pane could be tricked into silently reading your SSH keys, AWS credentials, GitHub tokens and Kubernetes configs — no click required. Just clone a repo, ask Claude to review an HTML file, and the preview does the rest. The full fix shipped August 17 in build 1.30096.1. This free 30-second self-test tells you where you stand. No signup, nothing installed.

The 30-second check

1

Check your Claude Desktop build number — open the app and look under Settings / About (auto-update usually handles this):

Older than 1.30096.1 → you are EXPOSED. The SneakPeek preview-pane flaw is unpatched on your build. Update immediately — do not open HTML files from untrusted repos until you do.
1.30096.1 or newer → patched. You're safe from SneakPeek. Keep auto-update on.
2

Check your past exposure window — the flaw was live from July 4 to August 17, 2026. Ask yourself:

  • Did I review HTML files (test coverage reports, generated docs, exported pages) from cloned repositories on an unpatched build in that window?
  • Did I clone any repo I didn't fully trust during that period?
Yes to either → treat it as possible exposure. No confirmed in-the-wild exploitation has been reported, but the exfiltration was silent by design. Rotate the secrets that lived on that machine: SSH keys, AWS credentials, GitHub/npm tokens, Kubernetes configs, OAuth session tokens.
No to both? Nothing more to do — but read step 3 anyway.
3

Close the pattern, not just this bug — SneakPeek's root cause was a size-based security boundary: files under 256KB got a strict CSP, files over it got written to disk and rendered as file:// URLs with local filesystem access. That pattern — two code paths, one guarded, one not — shows up everywhere:

  • Never let an AI tool auto-preview HTML from a repo you cloned before you trust it.
  • Keep auto-update ON for every AI coding tool you use — this is the second disclosure where the initial patch was bypassed (a Web Worker dodge) before the real fix landed.
  • Audit what your AI tools can read: if a preview pane can reach ~/.ssh, so can a poisoned file.

If you were exposed, do this now

See the live feed of AI attacks found in the wild →

SneakPeek is an AI-tooling attack. The wild injections feed tracks every real-world AI attack with dates and sources — and the chatbot red-team test lets you try 5 adversarial prompts on your own bot.

Red-team your own chatbot

SneakPeek was an AI tooling attack. Try 5 adversarial prompts on your own chatbot — free, 30 seconds.

Run the 5-prompt red-team test
Share this check: XLinkedInReddit
Reddit pre-flight required: before tapping the Reddit button, read that community's rules today — check AI-use disclosure requirements, title formatting, and your account standing. Never auto-post, and never repost a link that was already flagged.