Safety
Deployment-Time Memorization in Foundation-Model Agents
A new study examines how AI agents with persistent memory create privacy-utility tradeoffs in real-world deployments, finding that aggressive summarization can reduce data extraction risks by up to 76% while preserving personalization. The research reveals that even when information is "deleted," derived copies often remain recoverable in memory systems.
Read full story at arXiv →V: · A: · D:
Related
Safety
Predicting model behavior before release by simulating deployment
OpenAI has introduced a method called Deployment Simulation that uses real conversation data to anticipate how a model w...
Safety
Critical Copilot vulnerability allowed hackers to steal 2FA code from users
A now-patched vulnerability in Microsoft Copilot, dubbed SearchLeak, allowed attackers to exfiltrate two-factor authenti...
Safety
KPMG pulls report on AI usage due to apparent hallucinations
KPMG has withdrawn a research report about AI usage after discovering apparent hallucinations in the AI-generated conten...