Back to Intelligence

Anthropic Claude Cowork VM Escape: macOS Detection and Hardening Guide

SA
Security Arsenal Team
July 23, 2026
5 min read

Introduction

Security researchers at Accomplish AI have disclosed a critical security flaw in Anthropic's "Claude Cowork" application that fundamentally breaks the isolation model between AI agents and the host operating system. This vulnerability allows an AI agent running within a contained Linux virtual machine (VM) to escape its sandbox and read or write files anywhere on the macOS host filesystem.

With an estimated 500,000 macOS users potentially affected, this is not a theoretical edge case—it is a significant breach of trust and integrity. For defenders, the implication is stark: granting an AI agent access to code or documents via the Claude interface could inadvertently grant it full access to sensitive user data, SSH keys, credential files, and system configurations. We need to treat this specific AI integration with the same scrutiny we apply to untrusted binaries.

Technical Analysis

  • Affected Products: Anthropic Claude Cowork for macOS.
  • Affected Platform: macOS (Intel and Apple Silicon) hosting the Claude Cowork Linux VM environment.
  • Vulnerability Class: Sandbox Escape / VM Escape.
  • Underlying Mechanism: The application utilizes a Linux VM to execute agent tasks. The vulnerability lies in how the VM interacts with the macOS host filesystem, likely via shared folder mechanisms or improper hypervisor isolation. The flaw allows the agent to traverse outside the designated mount point.
  • Impact: Confidentiality breach (reading any file on the Mac) and Integrity breach (writing/modifying any file on the Mac).
  • Exploitation Status: Detailed vulnerability disclosed by researchers. While no specific CVE identifier was provided in the disclosure, the PoC confirms the ability to break out of the Linux confinement.

Detection & Response

Detecting a VM escape requires monitoring for abnormal process interactions between the application wrapper and the host filesystem. We are hunting for the Claude application spawning shells or accessing sensitive directories that should be outside its scope.

SIGMA Rules

YAML
---
title: Claude Cowork Suspicious Shell Spawning on macOS
id: 8f4a2c1e-9b3d-4a5e-8f6d-1c2b3a4c5d6e
status: experimental
description: Detects the Claude Cowork application spawning a shell (bash/zsh), indicating potential command execution on the macOS host.
references:
  - https://thehackernews.com/2026/07/claude-cowork-flaw-could-let-ai-agent.html
author: Security Arsenal
date: 2026/07/15
tags:
  - attack.execution
  - attack.t1059.004
logsource:
  category: process_creation
  product: macos
detection:
  selection:
    ParentImage|contains: 'Claude'
    Image|endswith:
      - '/bash'
      - '/zsh'
      - '/sh'
  condition: selection
falsepositives:
  - Legitimate developer debugging (rare)
level: high
---
title: Claude Cowork Accessing Sensitive User Directories
id: 9e5b3d2f-0c4e-5b6f-9g7e-2d3c4b5e6f7g
status: experimental
description: Detects Claude Cowork processes accessing sensitive directories like .ssh or .aws, indicative of a VM escape or data exfiltration attempt.
references:
  - https://thehackernews.com/2026/07/claude-cowork-flaw-could-let-ai-agent.html
author: Security Arsenal
date: 2026/07/15
tags:
  - attack.collection
  - attack.t1005
logsource:
  category: file_access
  product: macos
detection:
  selection:
    Image|contains: 'Claude'
    TargetFilename|contains:
      - '/.ssh/'
      - '/.aws/'
      - '/.kube/'
  condition: selection
falsepositives:
  - User explicitly granting access to these folders for legitimate AI tasks
level: high

KQL (Microsoft Sentinel / Defender)

KQL — Microsoft Sentinel / Defender
// Hunt for Claude Cowork processes interacting with sensitive file paths
DeviceProcessEvents
| where Timestamp > ago(1d)
| where InitiatingProcessFileName contains "Claude" 
| where ProcessCommandLine has "/.ssh/" 
   or ProcessCommandLine has "/.aws/" 
   or ProcessCommandLine has "/etc/" 
| project Timestamp, DeviceName, AccountName, InitiatingProcessFileName, ProcessCommandLine, FolderPath
| order by Timestamp desc

Velociraptor VQL

VQL — Velociraptor
-- Hunt for Claude processes accessing files outside of standard user directories
SELECT Pid, Name, CommandLine, Exe, Username
FROM pslist()
WHERE Name =~ 'Claude'
   AND (CommandLine =~ '/etc/' OR CommandLine =~ '/.ssh/' OR CommandLine =~ '/.aws/')

Remediation Script (Bash)

Bash / Shell
#!/bin/bash

# Claude Cowork Hardening Script for macOS
# This script checks for the presence of the app and attempts to quarantine it if an update is not available.

APP_NAME="Claude.app"
APP_PATH="/Applications/$APP_NAME"
QUARANTINE_ATTR="com.apple.quarantine"

if [ -d "$APP_PATH" ]; then
    echo "[+] Claude Cowork found at $APP_PATH"
    
    # Check if the app is currently running
    if pgrep -f "$APP_NAME" > /dev/null; then
        echo "[!] WARNING: Claude Cowork is currently running. Please quit the application before proceeding."
        exit 1
    fi

    echo "[*] Checking for updates via Homebrew (Cask) or manual intervention required..."
    # Note: Automated updating is discouraged to prevent version drift. 
    # Administrators should verify the latest patched version from Anthropic.

    echo "[!] ACTION REQUIRED: Verify patch status for the VM Escape vulnerability."
    echo "[!] If no patch is available, restrict Full Disk Access in System Settings."

    # Optional: Quarantine the app to prevent execution until patched (Uncomment below to enforce)
    # xattr -w "$QUARANTINE_ATTR" "0181;$(date +%Y-%m-%d);Security Arsenal; Anthropic Claude Cowork Quarantine" "$APP_PATH"
    # echo "[+] App has been quarantined."

else
    echo "[-] Claude Cowork not found in /Applications."
fi

Remediation

  1. Immediate Patching: Check for updates to the Claude Cowork application immediately. Anthropic is expected to release a patch restricting the filesystem mount scope. Ensure all macOS endpoints are updated to the latest build where this isolation is reinforced.
  2. Restrict File Permissions: In macOS System Settings > Privacy & Security, review permissions granted to Claude Cowork. Revoke "Full Disk Access" and only grant access to specific, non-sensitive folders (e.g., a dedicated ~/Projects folder) if the agent requires file I/O.
  3. Network Segmentation: If the AI agent performs network calls, ensure egress filtering is applied to prevent data exfiltration if the escape occurs.
  4. User Awareness: Alert the user base (~500k users) that "Code Interpreter" or "File Analysis" features in AI tools should be treated with the same caution as running untrusted scripts.
  5. Vendor Advisory: Monitor the official Anthropic security advisory channel for the specific patched version number addressing this VM escape vector.

Related Resources

Security Arsenal Penetration Testing Services AlertMonitor Platform Book a SOC Assessment vulnerability-management Intel Hub

criticalzero-daycvepatch-tuesdayexploitvulnerability-disclosureanthropicclaude-coworkvm-escapemacos

Is your security operations ready?

Get a free SOC assessment or see how AlertMonitor cuts through alert noise with automated triage.