From 9bb48ec6cf6a4d11b75f59ba3f45b6e6f9815ad1 Mon Sep 17 00:00:00 2001 From: reacher-z Date: Tue, 28 Jul 2026 16:48:58 +0800 Subject: [PATCH] Add ClawBench as complementary browser benchmark --- resources/benchmarks.md | 11 +++++++++++ 1 file changed, 11 insertions(+) diff --git a/resources/benchmarks.md b/resources/benchmarks.md index 1af2324..9283c32 100644 --- a/resources/benchmarks.md +++ b/resources/benchmarks.md @@ -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: , , , and . +- 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.