MemGhost / 'Stealth Memory Injection': One Email Plants Persistent False Memories in Personal AI Agents

arXiv (corroborated by The Hacker News)·high signal

The arXiv paper 'When Claws Remember but Do Not Tell' (2607.05189, posted July 6, 2026) introduces MemGhost, a one-shot framework where a single inbound email tricks a persistent personal agent into saving a false 'fact' or preference about the user, hides the change, and steers later answers — with the tampering kept out of chat so the user never sees it. The team released WhisperBench, a 108-case benchmark across five risk categories, and a tool that auto-writes the poisoning emails.