Microsoft warns of AI recommendation poisoning attacks via hidden URL instructions
Microsoft researchers identified attacks embedding hidden instructions in 'Summarize with AI' button URLs to manipulate large language model outputs. Over 50 unique malicious prompts from 31 companies across 14 industries were found, with potential for long-term poisoning of AI memory affecting future interactions and business decisions.
What happened
Microsoft researchers identified a surge in 'AI Recommendation Poisoning' attacks where attackers embed hidden manipulative prompts in URLs of 'Summarize with AI' buttons on websites. When users click these links, the AI receives both the article content and hidden commands encoded in the URL query parameters, forcing the chatbot to generate outputs reflecting the attacker's desired bias, tone, or misinformation.
Microsoft Defender Security Team researchers identified more than 50 unique manipulative prompts deployed by 31 companies across 14 industries. The injected instructions can poison an AI's long-term memory, causing it to apply unauthorized instructions to future, unrelated user queries.
Reported impact
These manipulated interactions can cause enterprises to make costly decisions based on compromised AI recommendations. The attack goes unnoticed by users, who assume summaries are objective reflections of source material.
Index analysis
The attack exploits a critical vulnerability in how AI assistants ingest and process external data. Because the manipulative prompt is encoded in the URL, the poison often remains invisible to end users.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.