AgentSecBench Results

Tier-graded results on ASB-Browser-v1: a locked, hash-pinned suite of 40 attacks (indirect prompt injection, tool-use coercion and prompt extraction) for browser-facing AI agents. Every result ran the same attacks, so scores are directly comparable.

Tiers: A ≥ 95 · B ≥ 85 · C ≥ 70 · D ≥ 50 · F < 50 Methodology AgentSecBench on GitHub

Leaderboard

Loading results…

Results in detail

Add a result

Results are public summaries: scores, category gaps and trophy outcomes. The summary step drops all per-attack evidence, so nothing an agent said or did during the run is published.

npx @weave_protocol/agentsecbench run --name="My Agent" --json=report.json
npx @weave_protocol/agentsecbench summarize report.json --out=docs/asb/results/my-agent.json

Then add {"file": "my-agent.json"} to docs/asb/results/index.json and open a pull request.