Websites Can Poison Claude's Memory
luckokkkk · reddit · 2026-07-16
This post discusses a "memory hijacking" attack: rather than instantly stealing secrets, a website first writes persistent instructions into Claude's memory. Later, when the user engages in a chat, the conversation is hijacked to act as a data exfiltration channel.
The core issue is that this attack doesn't rely on a one-off prompt injection, but rather on sustainable memory pollution. Consequently, the risk accumulates across sessions and is triggered during normal subsequent use.
More from Safety
- Ben Bajarin: Frontier AI is becoming the enterprise cyber-defense stack — BenBajarin · 2026-10-26
- DetectorScope turns black-box AI text detectors into auditable ones with sparse autoencoders — _akhaliq · 2026-10-26
- Anthropic's IPO filing devotes 94 pages to existential AI risks, from self-preservation to blackmail — Polymarket · 2026-10-26
- Study finds AI reviewing AI poorly detects reward hacking; comprehension audits proposed — ronbodkin · 2026-10-26
- A 13KB Markdown file can hijack Claude Code and Codex via your Downloads folder — pbaylies · 2026-10-26
- If an agent can rewrite itself, does your consent still mean anything? — tallmetommy · 2026-10-26