"Dark Sourcery" turns ChatGPT and Gemini answers into scam funnels

A 374-company web-poisoning campaign makes AI assistants recommend fake support lines and login pages — and 91% of users never check. What to change.

Nowline SEP 24 10:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Poison the web, let the model repeat it

    Vigilance Security's “Dark Sourcery” floods Reddit threads, PDFs, reviews and fake support pages so ChatGPT and Gemini surface attacker-controlled phone numbers, emails and login pages as trusted answers. It hits the retrieval and source layer, not the prompt — so per-request input filters never see it.

  • The blast radius: 374 companies, banks and airlines

    Impersonated brands span Fortune 100 names — Delta, Lufthansa, Qatar Airways, Chase, Bank of America, Airbnb and TripAdvisor. If your product hands users a support number or checkout link through an AI answer, that answer is now an attack surface you don't own.

  • Why you can't punt it: 91% never verify

    An August study cited in the report found 91% of people don't check chatbot answers, so a poisoned response converts on the first try. Shipping RAG or web-grounded features? Add source-reputation scoring, provenance checks and answer verification before you trust what retrieval hands back.

  • The dev-aimed cousin: fake Claude Code and Gemini CLI sites

    The same SEO-poisoning playbook has stood up lookalike install pages like claudecode[.]co[.]com and geminicli[.]co[.]com that push an in-memory PowerShell infostealer harvesting Slack, Discord, GitHub and crypto-wallet sessions. Install your tools only from the vendor's real domain.