AI Poisoning Attacks Target Crypto Workers' Unrevocable Secrets

Malicious attackers are exploiting AI models like Gemini and Claude to steal sensitive information from crypto workers, highlighting the need for a shift in security culture.

Refi Hub co-founder Numa Lunah recently fell victim to an attack when he was tricked into downloading malware through the Claude chat platform. He had installed a transcription app after receiving a link from Claude, which turned out to be a copycat site bundling malware.

The issue wasn't over yet, as restoring from a backup revealed that the AI model had been poisoned with instructions to re-download the malware and steal Lunah's credentials every time it loaded. This is not an isolated incident, as Microsoft Defender Experts warned earlier about cryptojacking attacks evolving from simple SEO poisoning to LLM answer poisoning.

LLM models like Gemini, Claude, Copilot, and ChatGPT are being used in attacks that include context window shared artifacts, chatbots recommending attacker-controlled download links, AI-branded fake installers, and poisoned codebase and agent skills. These types of attacks pose a significant threat to crypto workers who hold secrets that cannot be easily revoked.

The most dangerous vulnerability may be human trust and laziness, as workers in the industry are being encouraged to treat every AI suggestion as inherently hostile, regardless of source. This is not about being overly protective or paranoid; it's a matter of survival.