audr
github

audr.dev · live: CVE-2026-41386 published 2d ago

OpenClaw is on every laptop your team owns.
CVE-2026-41386 means setup codes mint admin-capable devices.
Audr finds where setup codes outgrow their role.

audr is an offline, single-binary scanner for AI agent configs — Claude Code, Cursor, Codex CLI, Windsurf, MCP servers. Run it on a laptop, get an HTML report. CISO-shaped, dev-deployable. 

curl -fsSL https://audr.dev/install.sh | sh
View sample report →

Or read the README · /security

in-browser scan · no install · paste anything

Same engine the binary ships. Compiled to WASM. Runs in your tab.

audr scan · in-browserready
Paste your own config above or pick a sample tab. Nothing is uploaded — the scanner runs entirely in this tab.
~/.mcp.json

Your config never leaves your browser. View source: github.com/harshmaur/audr-web. WASM blob SHA-256: f1ef9e068130…

recent advisories · audr v0.3.1 detects every one

Five CVEs your engineering org probably hasn't checked yet.

CVE-2026-41386

OpenClaw · OpenClaw

Bootstrap setup codes are not role-bound during first-use pairing.

✓ rule: openclaw-unbound-bootstrap-setup-code

CVE-2026-41349

OpenClaw · OpenClaw

config.patch can silently disable execution approval.

✓ rule: openclaw-config-patch-consent-bypass

CVE-2025-59536

Anthropic · Claude Code

Settings hooks execute arbitrary shell on PreToolUse / Stop events.

✓ rule: claude-hook-shell-rce

CVE-2026-25253

OpenClaw · OpenClaw

MCP credentials stored plaintext in user-readable config.

✓ rule: mcp-plaintext-api-key

CVE-2026-39861

Anthropic · Claude Code

Symlink in workspace allows sandbox escape on read.

✓ rule: claude-skip-permission-prompt

what a real scan looks like

Forensic HTML report. Redacted excerpt — 200 dev laptops, real findings.

audr emits an HTML report you can forward up the chain. Each finding has a one-line attacker-gets, a file:line citation, and lands on the Attack Chain it belongs to.

audr scan · redacted excerpt · 200 dev laptops

200

laptops scanned

14

critical findings

5

attack chains

4.2s

scan time

Open the full sample report →