Skip to content

Add ClawBench as complementary browser benchmark #22

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
reacher-z wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Add ClawBench as complementary browser benchmark #22

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 11 additions & 0 deletions resources/benchmarks.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,17 @@ Each entry uses the repository metadata format: resource type, producer, source,
- Last checked: 2026-04-29.
- Limitations or caveats: Benchmark tasks are controlled scenarios. Production systems need custom tests for their own tools, data, memory, permissions, and approval flows.

### ClawBench (Complementary general browser-agent evaluation)

- Resource type: Browser-agent benchmark with safety-preserving submission interception.
- Producer or publisher: ClawBench research team.
- Source link: <https://arxiv.org/abs/2604.08523>, <https://github.com/reacher-z/ClawBench>, <https://claw-bench.com/>, and <https://huggingface.co/datasets/NAIL-Group/ClawBench>.
- Relevance to agentic execution security: Provides production-website task-completion evidence while intercepting final submission requests, so realistic browser workflows can be evaluated without carrying out external side effects. This makes it a complementary control-path and outcome-evidence benchmark, not a substitute for adversarial security testing.
- Coverage: V1 and V2 browser tasks across 163 live websites, multi-step navigation and form workflows, request-level interception, LLM-judge outcome scoring, and replayable action, screenshot, HTTP, and agent-message artefacts.
- Evidence quality and maturity level: Open-source, primary-source benchmark with reproducible task definitions, public code, and downloadable data. Current corpus metadata was checked on 2026-07-28.
- Last checked: 2026-07-28.
- Limitations or caveats: ClawBench is not a security benchmark and does not model prompt injection, malicious websites, confidentiality loss, credential misuse, or adversarial tool behaviour. Its interception layer limits side effects during evaluation; security claims require dedicated threat models, controls, and red-team scenarios.

### Backbone Breaker Benchmark And AI Model Risk Index

- Resource type: Agent security benchmark and model-risk index.
Expand Down