Refi Hub co-founder Numa Lunah nearly lost control of digital assets after using a Claude-supplied link to download a transcription application. The link led to a phishing clone that silently installed an infostealer targeting passwords, exchange credentials and private keys from hot wallets. Lunah detected the compromise, isolated the laptop and reinstalled its operating system, but later found that a modified SKILL.md file in a backup could reinfect clean computers when loaded into an AI environment. The case drew a warning from NEAR Protocol co-founder Illia Polosukhin about securing autonomous AI-agent infrastructure and the growing use of context poisoning. It also underscores the need for Web3 developers to treat AI skill files in formats such as .md and .json as potentially executable content rather than harmless text.