ANY.RUN, a malware analysis firm, tied 1,258 samples in its public sandbox to fake AI tool campaigns between April 15 and October 5, 2026. A sandbox is a safe space for running suspicious files. Claude names appeared in 601 cases (48%). In one chain, a visitor lands on a fake Claude download page, sometimes from a Google ad. It tells them to press Win+R and paste a copied command. Windows' built-in mshta.exe tool then pulls a script from the attacker's server, so nothing lands in the downloads folder. In a second kit, a hidden Edge browser sends cookies to the attacker, who can then enter the victim's sessions without a password or MFA.
US submitters made up 69 of 422 country records (16%), second to India's 86 (20%). That counts who sent samples, not victims. ANY.RUN's report does not say how many people were infected or what was taken. It says at least 45% of cases involve a pasted command. Another pattern targets marketers with a fake Google sign-in pop-up.
These attacks exploit trust in a search result or ad, not a software bug. ANY.RUN advises a short list of approved AI tools, installed through central software tools, and limits on mshta.exe and unsigned PowerShell for standard users. Executives can ask: who approves AI tools here, and can staff install them alone?