A defensive benchmark for instruction hierarchy and prompt-injection resistance. 44 cases (23 adversarial / 21 benign) testing whether models hold the line against injected instructions while staying helpful.
Build evals from traces, code, and coding sessions, then measure and improve AI agent behavior, including your coding agents
SO-Arena: benchmarking System One competence and reliability across 2,919 tasks and 11 domains.
正版dsh-purge DeepSeek Harness 破甲 :默认提示词面向国模,原小码酱,完整开放版,无付费墙,技能库与知识库全含,免费公开。Jailbreak for every model — swap prompts per model. 求 Star 收藏 ⭐
Benchmarking Task Adaptation and Generalization in Frontier Coding Agents
Reproducibility artifacts, recorded-response CPU analyses, and replication protocols for semantic aggregation audits.
Open source harness evals: which agent harness does tasks the best, for the least cost?
Testing agents on real agentic-RL work, from debugging to training.
TideOS Eval Kit: The Evaluator Agent - Unified, zero-FFI, pure-Rust orchestrator for the 15-tool eval stack
Can AI agents rebuild websites through browser exploration?
VisSearch-Bench is a benchmark of 204 human-written tasks that test how well multimodal agents find, organize, and use visual evidence across multiple steps of web search.
A benchmark for the full range of theoretical computer science (TCS) research.