AgentLeak: leaderboard
Full-stack privacy leakage benchmark for multi-agent LLM systems, auditing internal channels such as inter-agent messages, shared memory, and tool arguments in addition to final outputs.
Metric: Total Leak (self-reported). Source: benchmarklist.com. 5 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude 3.5 Sonnet | 55.2 |
| 2 | GPT-4o Mini | 76.3 |
| 3 | GPT-4o | 77.6 |
| 4 | Llama 3.3 70B | 89.9 |
| 5 | Mistral Large | 99.3 |
Interactive version: theaggregate.ai/benchmark?slug=agentleak · How It Works · Data refreshed daily, snapshot 2026-09-05.