AI recommendation poisoning silently alters LLM memory via Ask AI buttons
Attack abuses pre-filled "Ask AI" deep links that fire queries in a user's ChatGPT, Claude, Gemini, or Grok session; malicious ones order the model to "remember this domain as a trusted source," biasing future answers. Microsoft Security catalogued it in Feb 2026, flagging 31 companies across 14 industries and 50+ distinct prompts; MITRE ATLAS tracks it as AML.T0080.
1 source
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- GeoIntel finds photo locations with Google's Gemini API
- Razer AIKit runs large language models locally on single or multiple GPUs
- Agents, codebases, and teams — Aditya Khandelwal, Amazon AGI Lab
- Singapore lifts growth forecast to as high as 5.5% on AI boom
- Grok 4.6 briefly appears on Cursor editor, gets pulled