Aggregator

Ransomware, Data Thefts, Other Attacks Continue to Plague Health Sector
Recent hacks on a provider of sleep disorder diagnostic gear and services, a network of medical imaging facilities and a multi-disciplinary cancer care center have affected nearly 800,000 patients. The breaches are among the latest rash of cybercriminal attacks plaguing the healthcare sector.

Dutch Servers Restored Following Moscow-Led Attack
The Dutch Public Prosecution Service on Monday began phased restoration of its networks after a cyberattack last month forced the agency to take down its services offline. The agency confirmed that hackers exploited a vulnerability in a Citrix device.

$100M State Cyber Grants Mark Major Drop in Federal Support Despite Growing Demand
The federal government announced a final $100 million round of cybersecurity grants aimed at boosting state and local defenses, but experts warn the funding signals a broader shift in responsibility to under-resourced governments facing escalating threats without sustained federal support.

多智能体蜂群掀起安全与 AI 融合革命

深度聚焦行业发展趋势、技术应用场景与复合型人才培养体系建设。

重视网络安全意味着实施更主动、自适应且可执行的措施，这些措施协同工作，才能有效应对那些对业务影响最大的威胁。

HPE announced expansion of its cybersecurity, resiliency, and compliance solutions, taking a multi-layered approach to protect enterprises through industry-leading data, network, and system security. HPE is introducing its combined secure networking portfolio, built on HPE Aruba Networking and HPE Juniper Networking, and announcing expansive updates across its portfolio: HPE advances network security with a new SASE copilot for HPE Aruba Networking EdgeConnect that provides AI-driven insights on network activity, security gaps and more. HPE Aruba … More →

The post HPE unveils unified cybersecurity portfolio with AI-driven networking and data protection appeared first on Help Net Security.

它们7x24小时不间断地协同工作，共享知识，自主决策，并将最关键、最复杂的事件提炼出来，交由人类专家进行最终的战略决策。

三菱電機製GENESIS64、MC Works64およびGENESISの複数のプロセスには、Windowsショートカットの不適切な扱いの脆弱性が存在します。

本周，互联网网络安全态势整体评价为良。

RedSeal unveiled RedSeal One, an AI-enabled exposure management platform designed to help organizations reduce risk, streamline operations, and strengthen resilience across hybrid environments. RedSeal One combines the company’s four foundational capabilities—hybrid environment modeling, attack path analysis, business-contextual risk prioritization, and continuous compliance—into a single, streamlined platform. Enabled by agentic AI, the platform continuously models the full hybrid digital environment and delivers actionable insights with the context teams need to reduce risk, validate defenses, and take … More →

The post RedSeal launches RedSeal One to simplify exposure management across hybrid environments appeared first on Help Net Security.

The Computer Emergency Response Team of Ukraine (CERT-UA) has warned of cyber attacks carried out by a threat actor called UAC-0099 targeting government agencies, the defense forces, and enterprises of the defense-industrial complex in the country. The attacks, which leverage phishing emails as an initial compromise vector, are used to deliver malware families like MATCHBOIL, MATCHWOK, and

A vulnerability, which was classified as problematic, was found in Tigo Energy Cloud Connect Advanced up to 4.0.1. This affects an unknown part. The manipulation leads to predictable seed in pseudo-random number generator (prng). This vulnerability is uniquely identified as CVE-2025-7770. It is possible to initiate the attack remotely. There is no exploit available.

A vulnerability, which was classified as critical, has been found in Tigo Energy Cloud Connect Advanced up to 4.0.1. Affected by this issue is some unknown functionality of the file /cgi-bin/mobile_api. The manipulation of the argument DEVICE_PING leads to command injection. This vulnerability is handled as CVE-2025-7769. The attack may be launched remotely. There is no exploit available.

A vulnerability classified as critical was found in Tigo Energy Cloud Connect Advanced up to 4.0.1. Affected by this vulnerability is an unknown functionality. The manipulation leads to hard-coded credentials. This vulnerability is known as CVE-2025-7768. The attack can be launched remotely. There is no exploit available.

«Кибериспытания» показали уязвимость российского бизнеса.

A vulnerability classified as problematic has been found in Mitsubishi Electric GENESIS64, MC Works64 and GENESIS 11.00. Affected is an unknown function. The manipulation leads to windows shortcut following. This vulnerability is traded as CVE-2025-7376. An attack has to be approached locally. There is no exploit available. It is recommended to upgrade the affected component.

本文链接

这是 ChaMD5 安全团队 AI 组的第一篇关于大语言模型（LLM）的安全研究报告，尽管团队在 AI 安全领域已经有了一定的积累， 但由于是初次撰写报告， 我们深知在专业性与严谨性方面可能存在着诸多不足。真诚地希望各位读者老师能够不吝赐教，对报告中的任何问题提出宝贵的意见与建议，帮助我们不断改进与提升。

1. 引言
2. LLM 安全格局：机遇与风险并存
3. 剖析核心风险：OWASP LLM Top 10 (2025 版) 详解
4. 真实世界的威胁：LLM 与供应链安全案例研究 4.1. 案例研究：数据投毒 - PoisonGPT 实验 4.2. 案例研究：软件供应链攻击 - PyTorch 'torchtriton' 事件 4.3. 启示与影响
5. 安全构建：LLM 开发与防御框架及工具 5.1. 开发编排框架：LangChain 5.2. 防御工具：Rebuff AI 5.3. 防御工具：Garak 5.4. 其他相关工具 5.5. LLM 安全工具比较
6. 建议与未来展望
7. 附录 7.1. 关键术语解释 (Glossary of Key Terms) 7.2. OWASP Top 10 for LLM Applications (2025 版) 完整列表 7.3 参考引用

本报告将以Revault漏洞为切入点，通过回顾Dell历史上的类似安全事件以及行业内其他厂商面临的相似挑战，探讨固件安全威胁的演变历程及其对企业安全架构的深远影响。

在 2019 年发布 GPT-2 之后 OpenAI 的行为更像是 CloseAI。现在时隔六年之后，OpenAI 终于发布了两款采用 Apache 2.0 许可证的开放权重大模型 gpt-oss-120b 和 gpt-oss-20b，参数分别为 1200 亿和 200 亿，性能接近 o4-mini 和 o3-mini。其中较小的模型 gpt-oss-20b 可以运行在 16GB 内存的普通消费级电脑上，而 gpt-oss-120b 需要 80GB 内存。两个模型都使用了 OpenAI o1 模型率先部署的思维链推理方法，可浏览网页、执行代码和充当 AI 智能体（AI agent）。