Add new entry for AgentLeak benchmark in README

[yagobski]

Summary

This PR adds AgentLeak as entry #19 in the Papers table, under the Agent Privacy category.

Paper: AgentLeak: A Full-Stack Benchmark for Privacy Leakage in Multi-Agent LLM Systems
arXiv: https://arxiv.org/abs/2602.11510
Code: https://github.com/Privatris/AgentLeak
Authors: Faouzi El Yagoubi, Godwin Badu-Marfo, Ranwa Al Mallah (Polytechnique Montreal)

Why This Paper Belongs Here

AgentLeak fills a gap in the existing list: it's the first benchmark to audit privacy leakage across all inter-agent communication channels in multi-agent LLM systems.

Key findings:

• Covers all 7 communication channels (system prompts, user inputs, tool calls, tool outputs, inter-agent messages, memory, external APIs)
• Inter-agent messages exhibit 68.8% leakage rate — the highest of all channels
• Output-only auditing (the standard approach) misses 41.7% of privacy violations
• 1,000 scenarios across healthcare, finance, legal, and corporate domains

This complements existing entries on prompt injection, data extraction, and general LLM security.