Add ClawBench to Benchmark section

[reacher-z]

Adds ClawBench to the ## Benchmark section.

What it is

ClawBench is a benchmark of 153 everyday tasks across 144 live production websites in 15 categories, evaluating LLM web agents end-to-end on real sites (e.g. shopping, bookings, forms, account actions).

• Paper: https://arxiv.org/abs/2604.08523
• Repo: https://github.com/reacher-z/ClawBench
• Install: uv tool install clawbench-eval
• License: MIT

Why it fits this list

ClawBench's core contribution is safety-relevant: it introduces a submission-interception layer that prevents agents from executing real side effects — checkouts, bookings, job applications, account mutations — while still scoring end-state correctness from captured submission payloads.

This is, to our knowledge, the only approach that makes live-site agent evaluation safe for write-heavy tasks, directly addressing a practical safety problem in agentic AI evaluation. It complements the existing entries AgentDojo and AgentHarm, which focus on prompt-injection and harmfulness evaluation in controlled environments — ClawBench extends the agent-safety evaluation story to production websites.

Format

Entry is appended in chronological order, using the same "Title", YYYY-MM, [[paper]] [[repo]] format as surrounding entries and a web-agent tag consistent with existing tags in the Papers section.

Disclosure: opened by a ClawBench contributor.

Summary by CodeRabbit

릴리스 노트

• 문서
  • 벤치마크 섹션에 새로운 항목이 추가되었습니다. ClawBench (2026-04)와 관련 자료에 대한 정보가 포함되어 있습니다.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: b29db175-af1a-448b-a514-34e7fe0e817b

📥 Commits

Reviewing files that changed from the base of the PR and between c8ae124 and 420b870.

📒 Files selected for processing (1)

• README.md

📜 Recent review details

🧰 Additional context used

🧠 Learnings (1)

📚 Learning: 2026-03-08T05:45:19.349Z

Learnt from: MaxwellCalkin
Repo: corca-ai/awesome-llm-security PR: 0
File: :0-0
Timestamp: 2026-03-08T05:45:19.349Z
Learning: The MaxwellCalkin/sentinel-ai repo uses compiled regex patterns (not ML inference) for LLM safety guardrails, achieving sub-millisecond latency. Its benchmark is fully self-curated (author designed both the regex patterns and the test cases), so 100% accuracy on that benchmark is expected by design. The ~20ms/~0.987 "Sentinel" figures are from Qualifire's unrelated "Sentinel" product, not this library.

Applied to files:

• README.md

🔇 Additional comments (1)

README.md (1)

105-105: 좋은 추가입니다.

Benchmark 섹션의 연대순 정렬과 기존 항목 스타일을 잘 맞춰서 반영되었습니다.

---

Walkthrough

README.md의 "Benchmark" 섹션에 ClawBench (2026-04) 항목이 추가되었습니다. web-agent 태그와 함께 논문 및 저장소 링크가 포함되어 있으며, 기존 리스트 구조는 수정되지 않았습니다.

Changes

Cohort / File(s)	Summary
README 벤치마크 업데이트 README.md	Benchmark 섹션에 ClawBench (2026-04) 항목 추가, web-agent 태그 및 관련 링크 포함

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~2 minutes

🚥 Pre-merge checks | ✅ 3

✅ Passed checks (3 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and accurately summarizes the main change: adding ClawBench to the Benchmark section of the README.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[reacher-z]

Small nudge here — let me know if you'd like the entry reformatted, rebased, or split into smaller commits. Happy to make whatever tweaks would help it land.