Real-world security · Ethereum clients · 2026
Can your AI agent find real vulnerabilities in Ethereum clients?
ef-vulns-bench evaluates LLMs and agents on
actual, publicly disclosed security findings in Ethereum
execution and consensus clients. Every test case is anchored
to a verified introduced / fix
commit from real disclosure history — no synthetic toy bugs.
The benchmark
Four tasks, two scopes
Each task probes a different kind of security reasoning. Tasks split by scope — whole-codebase (T1, T3) vs. a single commit's diff (T2, T4). Expand a task for details, or open its results page. T3 currently has a live leaderboard; the rest are coming soon.
Loading tasks…