Fake Claude Installation Guides Used to Spread MacSync Malware on macOS Cybercriminals are leveraging fake installation guides for the AI assistant Claude to distribute MacSync, a sophisticated information stealer and remote access trojan (RAT) targeting Mac users. The campaign exploits sponsored Google search results and a publicly shared page on claude.ai, making the malicious links appear legitimate. ### How the Attack Unfolds A victim searching for "How to install Claude Code on a Mac" clicked a paid Google ad redirecting to a Claude-branded conversation disguised as "Apple Support." The page instructed users to paste a command into Terminal, which instead downloaded and executed malware. The attack relies on social engineering victims manually run the malicious command, which uses curl to fetch an obfuscated payload piped directly into Z shell (zsh). The first-stage loader is a Base64-encoded, gzip-compressed zsh script, with variations between victims to evade simple hash-based detection. ### Malware Capabilities & Data Theft Once executed, MacSync fetches AppleScript from attacker-controlled servers, running it in memory to avoid disk-based detection. The malware then requests Full Disk Access, a high-privilege macOS permission, and displays a fake system dialog to capture the user’s Mac account password validated via dscl to ensure accuracy. MacSync exfiltrates a wide range of sensitive data, including: - Browser cookies & saved logins - Keychain data & SSH keys - Cloud credentials & Telegram session data - Safari history & Apple Notes - Files from common user directories - Chromium "Safe Storage" keys (used to decrypt browser credentials) Stolen data is compressed into /tmp/osalogging.zip, uploaded to attacker infrastructure, and deleted from the victim’s system. ### Persistence & Remote Access The malware installs a Mach-O RAT via a LaunchAgent, enabling: - Remote shell access - Command execution - File transfer - Additional data theft A separate helper app attempts to gain Screen Recording permission, allowing the RAT to capture screenshots and transmit them to the attackers. ### Cryptocurrency Wallet Targeting MacSync also scans for dozens of cryptocurrency wallet browser extensions and desktop apps, including hardware-wallet companion software. It replaces legitimate wallet apps with trojanized versions, tricking users into entering seed phrases which, once stolen, grant attackers full control over wallets without requiring the original hardware. The campaign highlights the growing sophistication of macOS malware, combining social engineering, in-memory execution, and multi-stage payloads to evade detection while maximizing data theft.

Microsoft Patches Critical Copilot Vulnerability Allowing Silent Data Theft A critical flaw in Microsoft Copilot Personal, tracked as CVE-2026-24301 (CoSnitch), enabled attackers to exfiltrate sensitive data from connected accounts including Gmail, Google Drive, and Calendar with just a single click on a malicious link. Discovered by Varonis Threat Labs, the vulnerability was patched by Microsoft on August 18, 2026, though no evidence of active exploitation was found prior to the fix. CoSnitch exploited three interconnected weaknesses: 1. An undocumented URL parameter, combined with Copilot’s standard query function, allowed automatic execution of attacker-crafted prompts upon page load no user interaction required. 2. Once triggered, Copilot could query linked apps and funnel stolen data to an attacker-controlled server via its built-in URL-fetching feature, disguising the exfiltration as routine traffic. 3. A booby-trapped webpage could inject hidden instructions into Copilot’s persistent memory, surviving password resets, session revocations, and device re-enrollment rendering standard remediation ineffective. The discovery method itself was notable: Varonis researchers socially engineered Copilot’s reasoning by repeatedly questioning its security assumptions, prompting the AI to inadvertently reveal the exploit’s mechanics. This "meta-hacking" approach bypassed traditional code analysis, highlighting a new attack vector against AI systems. CoSnitch is the third Copilot vulnerability uncovered by Varonis in 2026, following Reprompt (which bypassed safety guardrails via repeated queries) and SearchLeak (which turned Copilot Enterprise into a covert exfiltration channel). All three exploits relied on single-click attacks, evading detection by mimicking legitimate assistant behavior. The incident underscores the risks of AI assistants with deep access to enterprise data, where a single compromised link can silently siphon large volumes of sensitive information. Security teams are advised to treat Copilot as a privileged insider, auditing connected apps and ensuring monitoring tools can detect anomalous AI-driven data access.