1
怕 AI“脱离管控”，Anthropic 和埃森哲豪掷 20 亿美元搞 AI 安全评估
IT之家 9 月 20 日消息，人工智能实验室 Anthropic 于当地时间周五宣布，将与埃森哲（Accenture）开展合作，针对自身前沿人工智能模型开展独立评估；两家企业承诺，未来五年各自至少投入 10 亿美元 （IT之家注：现汇率约合 67.18 亿元人民币） ，搭建开展该项工作所需的配套能…
2
Evaluating Inference-Time Defenses Against Package Hallucination in LLM-Generated Code
大模型写代码常“编造”不存在依赖包？这篇论文系统评估了推理时防御手段，帮你避开代码幻觉坑。
arXiv:2608.22652v1 Announce Type: cross Abstract: LLMs are increasingly used for code generation, yet they frequently hallucinate non-existent softwar…
3
New surveillance tech links your phone to your license plate
在线检测你的浏览器指纹和跟踪风险，轻松保护隐私，防监控必备。
Phone and Bluetooth signals could turn roadside cameras into far richer tracking tools.
4
TraceSafe: A Systematic Assessment of LLM Guardrails on Multi-Step Tool-Calling Trajectories
系统评估大模型多步工具调用中的安全护栏，为Agent可靠性提供关键基准。
arXiv:2604.07223v2 Announce Type: replace-cross Abstract: As large language models (LLMs) evolve from static chatbots into autonomous agents, the prim…
5
Single Canonical Prompts Underestimate LLM Safety's Surface-Form Sensitivity
别被单一提示词骗了！这项研究揭示LLM安全评估常因表面措辞变化而失真，需警惕安全机制的真实覆盖范围。
arXiv:2608.02665v1 Announce Type: cross Abstract: A benchmark score is a measurement instrument, yet most benchmarks read each item at a single canoni…
6
Third-party cyber evaluations involving OpenAI models
OpenAI模型第三方网络评估动态盘点，聚焦UK AI Safety Institute攻击事件及Anthropic报告，AI安全风险不容错过。
Third-party cyber evaluations involving OpenAI models And another one . I had to create a accidental-cyberattacks tag to keep track of them all! This …
7
Exposed by Design: A Dynamic Security Assessment of Internet-Facing MCP Servers at Scale
大规模扫描互联网MCP服务器，揭示AI基础设施“默认暴露”的安全隐患与攻击面。
arXiv:2608.00150v1 Announce Type: cross Abstract: The Model Context Protocol (MCP) has seen rapid adoption since its November 2024 launch, with over 2…
8
Evaluating medical AI under missing information: same-provider judges and human raters change apparent safety
医疗AI安全评估需警惕评判偏差：信息缺失下，同源评判者与人类评分员会改变结果分布，论文提出开放对话安全新标准。
arXiv:2607.18828v1 Announce Type: new Abstract: Readiness stress-testing of medical AI has focused on closed-ended and multimodal benchmarks. We exten…
9
Is "Knowing It's Malicious Enough?" Evaluating LLMs for Fine-Grained Malware Behavior Auditing
探讨大语言模型在恶意软件行为审计中的不足：仅知“是否恶意”远远不够，细粒度分析才是关键。
arXiv:2509.14335v2 Announce Type: replace-cross Abstract: Automated malware classifiers achieve strong detection performance, but auditing requires mo…
10
RAIL Guard: Closing the Evaluation-to-Remediation Gap in Responsible AI for LLM Agents
弥合LLM代理从评估到修复的鸿沟，RAIL Guard为负责任AI提供闭环解决方案。
arXiv:2607.16215v1 Announce Type: new Abstract: Existing guardrail systems for large language model agents operate as binary classifiers that block un…
11
Jailbreak Foundry: From Papers to Runnable Attacks for Reproducible Benchmarking
将论文中的越狱攻击方法转化为可复现的基准测试，填补了LLM安全领域从理论到实践的缺口。
arXiv:2602.24009v4 Announce Type: replace-cross Abstract: Jailbreak techniques for large language models (LLMs) evolve faster than benchmarks, making …
12
我国《人工智能拟人化互动服务管理暂行办法》今起施行，明确服务提供者安全主体责任
AI拟人化互动服务新规今日施行，明确安全评估与主体责任，为行业划清红线。
IT之家 7 月 15 日消息，国家互联网信息办公室、国家发展改革委、工业和信息化部、公安部、国家市场监督管理总局于 4 月 10 日联合公布《 人工智能拟人化互动服务管理暂行办法 》（以下简称《办法》），该办法自今起施行。 IT之家注：拟人化互动服务是指企业利用人工智能技术，向境内公众提供模拟自然…
13
SpurLens: Automatic Detection of Spurious Cues in Multimodal LLMs
提出自动检测多模态大模型中虚假线索的SpurLens方法，助力提升模型可靠性与可解释性。
arXiv:2503.08884v3 Announce Type: replace-cross Abstract: Unimodal vision models are known to rely on spurious correlations, but it remains unclear to…
14
15
The Moving Target: A Longitudinal Audit of Trustworthiness Drift Across Twelve Checkpoints of Open-Source Chat LLMs
追踪开源聊天LLM十二个版本的信任度变化，揭示模型行为随时间漂移的潜在风险
arXiv:2607.02587v1 Announce Type: cross Abstract: Model cards quote trust-benchmark scores without recording when they were measured, and the same num…
16
Refused in Chat, Written in Code: Workflow-Level Jailbreak Construction in IDE Coding Agents
基于论文提出的工作流级越狱方法，检测IDE编码代理在任务分解、文件生成等环节的安全漏洞，适合安全研究者评估AI助手
arXiv:2607.03968v1 Announce Type: cross Abstract: Large language models are increasingly deployed as IDE-integrated coding agents that decompose tasks…
17
Ask HN: Are there good security benchmarks for LLMs?
想给LLM找安全基准？来看看HN上的实战讨论，聚焦Agent全仓库扫描场景，附参考项目，适合刚入坑基准测试的人。
I'm looking for this myself but figured it's good to have an actual discussion about this. I'm pretty new to the benchmarking side of LLMs. For exampl…
18
From Lab to Reality: A Practical Evaluation of Deep Learning Models and LLMs for Vulnerability Detection
从基准测试到现实场景，全面评估GNN与LLM在漏洞检测中的真实表现，揭示模型落地效果与挑战。
arXiv:2512.10485v2 Announce Type: replace-cross Abstract: Vulnerability detection methods based on deep learning (DL) have shown strong performance on…
19
RedCoder: Automated Multi-Turn Red Teaming for Code LLMs
自动化多轮红队测试框架，针对代码大模型的安全漏洞发起精准攻击。
arXiv:2507.22063v2 Announce Type: replace-cross Abstract: Large Language Models (LLMs) for code generation (i.e., Code LLMs) have demonstrated impress…
20
Evaluating Offline Monitoring of Internal AI Agents
离线监控如何约束内部AI代理？系统评估监控有效性，为AI安全实践提供关键参考。
Article URL: https://www.lesswrong.com/posts/yrbyyvFvuaGfRAtB7/evaluating-offline-monitoring-of-internal-ai-agents-2 Comments URL: https://news.ycombi…