AI系统黑盒API漏洞检测评估：KushoAI在真实业务逻辑缺陷检测中领先

一项针对七款AI系统在实时API中检测功能缺陷能力的黑盒评估报告发布。该评估使用KushoAI贡献的APIEval-20 v1.0开源基准，涵盖20个API场景、97个植入缺陷（分三个复杂度层级），各系统仅获得JSON模式和一个有效样本，需生成测试用例暴露缺陷，无源码和文档。结果显示，KushoAI在总体得分及各复杂度层级均排名第一，尤其在跨字段与业务逻辑缺陷检测上优势显著。报告还总结六大关键发现：看似完备的测试套件仍可能遗漏缺陷；简单模式级测试已成标配；提示工程可扩展广度但难以提升深度；复杂缺陷需组合有效字段形成无效状态；测试组合质量比数量更重要；一致性对CI/CD集成至关重要。该评估为AI驱动API测试工具选型提供了重要参考。 #AI #API测试 #黑盒测试 #漏洞检测 #KushoAI #LLM #软件开发 #基准测试

Solving the Worlds Hardest Problems with AI

Non-profit research consortium — free to read, forever The world needs better answers to its hardest questions We build peer-reviewed, openly licensed research on humanity’s most consequential problems. Anyone can contribute with AI assistance. Anyone can fund the team doing the work. Everything we publish is free for the world to use. Explore Problems How this works Why this exists Humanity’s biggest problems are chronically under-studied Pandemic risk, AI safety, antibiotic resistance, climate adaptation, mass poverty — the issues that will shape the century receive a fraction of the rigorous analysis we devote to incremental research. The gap isn’t intellect. It’s coordination: good minds working in silos, without a shared map, a shared rubric, or a permanent home for their work. The Consortium is that shared home — open, peer-review

ed, and built to outlast any single contributor. What's hot Surviving Birth: Closing the Maternal and Newborn Mortality Gap Maternal & Newborn Mortality 0 · 1d You don’t need a PhD or a big bank account — just the will to move something forward. Path 1 Contribute research Pick a problem. Draft a document. Use our built-in AI assistant to research, outline, and refine your thinking. Submit it for peer review and become part of the permanent knowledge base. Works with Claude, GPT, Gemini, and more — bring your own key No invite required — anyone can start a draft Accepted work is openly licensed and credited to you Learn how to contribute Path 2 Fund the work Don’t have time to write? Power the team that does. Your donation buys AI compute, peer-review time, and the infrastructure that keeps every published document free to read. 80% of every dollar goes directly to AI research & compute 15% keeps the platform open and running; 5% operations Transparent allocation — reported publicly See how funding is used What we stand for Principles we won’t bend on Open by default Every accepted document is published under a permissive license. No paywalls, no subscription, no gatekeeping. Evidence over opinion We use the ITN framework (Importance, Tractability, Neglectedness) and structured peer review to keep rigor high. Your keys, your data AI is bring-your-own-key. Your API credentials never leave your browser. We can’t read them; we don’t want to. 34 Problems ranked by ITN score 39 Documents in the library Free To read, fork, and cite — forever Where the work is needed Top priority problems Ranked by Importance, Tractability, and Neglectedness. 1 Global Health & Disease ITN 24 Preventable diseases kill millions annually. Malaria, tuberculosis, and neglected tropical diseases disproportionately affect the global poor despite cost-effective interventions existing. 2 Engineered Pandemics & Bioweapons ITN 23 Advances in synthetic biology lower the barrier to creating novel pathogens. A deliberately engineered pandemic could kill hundreds of millions and destabilize civilization. 3 Factory Farming & Animal Welfare ITN 23 Over 80 billion land animals are raised in factory farms annually under conditions of severe suffering. Alternative proteins and welfare reforms offer tractable near-term interventions. 4 Catastrophic AI Misuse ITN 22 AI systems used deliberately by state or non-state actors for mass casualties via cyberattacks on critical infrastructure, autonomous weapons, or supercharged bioweapon design. 5 Power-Seeking AI Systems ITN 21 Advanced AI systems may develop instrumental goals around self-preservation and resource acquisition, undermining human oversight and potentially leading to irreversible loss of control. Nothing about these problems solves itself. Whichever path fits your life right now, the Consortium is built so your effort compounds — every document, every dollar, every critique stays in the open for the next person to build on.

Sophos 发现黑客利用 Claude 设计出 70 种绕过杀毒软件的技巧

据安全公司 Sophos 披露，一个与勒索软件团伙有关的黑客团队在受控实验环境中接入 AI 代理及 Claude 大模型，批量生成了约 70 种绕过主流杀毒软件的技术。这些手法旨在规避安全产品的实时检测，显示生成式人工智能已被恶意用于自动化攻击工具的制造。Sophos 警告，此类利用 AI 批量开发绕过技术的行为可能降低网络犯罪门槛，对企业和个人终端安全构成新威胁。 #Sophos #网络安全 #黑客 #Claude #AI #勒索软件 #端点安全 #威胁情报 #人工智能滥用

台积电CEO坦言难以满足AI芯片需求

据路透社和彭博社报道，全球最大半导体制造商台积电正面临来自美国客户的巨大需求压力，即便其在美国的工厂正在扩建，仍难以完全满足订单。台积电CEO魏哲家在股东会后表示，客户需求极高，公司只能尽力支持，并强调正全力以赴避免成为供应链瓶颈。随着人工智能领域对先进芯片的需求激增，台积电的产能已接近极限，这凸显了全球半导体供应链的紧张态势。 #台积电 #半导体 #AI #芯片短缺 #供应链 #科技新闻