Benchmark RadarAI BENCHMARK PROFILE
WhisperBench
WhisperBench evaluates stealth memory injection attacks on persistent personal agents through a 108-case benchmark spanning five risk categories with fact and preference poisoning, using an IMAP/SMTP workflow and an email agent skill.
- Released
- 2026-07-06
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
The benchmark addresses the evaluation gap in assessing security of persistent memory in AI agents, providing a way to measure susceptibility to memory injection attacks and the effectiveness of defenses.
Motivation
Persistent personal agents combine long-term memory with access to users' external environments, enabling personalized foreground assistance and proactive background execution.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.