In the news
Utility Under Attack: Agent Memory Poisoning and the Limits of Content Screening and Provenance Ranking
arXiv cs.AI · Published · 3 min read
In 30 seconds
- What happened
- Researchers demonstrated that poisoning just 1.2% of an AI agent's memory with false statements drops accuracy from 85% to 30%, and content screening fails to catch it.
- Why it matters
- Matters for engineers building AI systems with persistent memory or retrieval-augmented generation that store and reuse information across sessions.
- Watch out
- Current defenses like content screening and provenance weighting either fail entirely or create unusable tradeoffs between blocking poison and allowing legitimate untrusted sources.
Listen to this summary
- agent
- retriever
- prompt
- prompt injection
- eval
The patterns behind this
Each one covers how the technique works, when it earns its cost, and where it breaks.
The Agent Architect
One pattern, one tradeoff, one production failure story. A short weekly briefing for people building agentic systems.
Weekly email, one-click unsubscribe. We only use your address to send the briefing.