AI-Linked Malware Threatens Crypto Security

Crypto professionals are being warned to verify any links or files shared through AI chat interfaces after a case where a user followed a download link from an AI chat window and was exposed to malicious code. Bitcoin.com News reported that Numa Lunah, co-founder of ReFi Hub, executed malware via a download link received in a Claude chat session. The report was published at 4:35 AM KST on the 30th. While Lunah stated no sensitive information was leaked, this claim has not been independently verified.
The key issue is not that Claude itself was hacked but that trusted AI interfaces can become vectors for malicious links. During recovery, Lunah discovered a 'poisoned SKILL.md' file meant to load malware every time the AI was used, stealing credentials. Microsoft's security team described this as an 'AI recommendation poisoning' attack in a February 2026 report, noting that clicking on seemingly legitimate AI service links can corrupt an AI assistant's memory and recommendations.
Microsoft identified over 50 prompt examples across 31 companies and 14 industries. Their method involves embedding prompts into links that appear to be normal AI service addresses but actually alter AI inputs or memories. Anthropic also lists prompt injection and data leaks as key risks of using skills from untrusted sources in Claude's help section, advising users to review file contents before activating.
Prompt injection tricks AI into processing external documents or links as if they were user commands. Skill files may look like simple documents but can trigger tool calls, file access, or external connections, making them a supply chain security concern. Huntress reported on July 21-22, 2026 that at least 29 organizations were infected with SectopRAT via malicious Claude artifacts hosted on claude.ai domains, with the links generating over 7,100 page views before removal.
BleepingComputer also covered a May 2026 case where Google ads combined with shared Claude chats distributed macOS malware. Users searching for 'Claude Mac download' were led through normal-looking ads and chat shares to execute terminal commands that collected browser credentials, cookies, and macOS keychain data.
The risk is heightened in crypto because professionals handle irreplaceable secrets like seed phrases, exchange API keys, hot wallet files, and deployer keys. The focus has shifted from AI answer accuracy to the broader security of all files and links processed by AI. Earlier reports emphasized designing workspace isolation, permission restrictions, audit logs, and human approval processes when connecting AI agents to code repositories and deployment systems.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: AI 링크 오염, 암호화폐 보안 새 공격면 됐다