|
|
|
|
@@ -58,7 +58,7 @@ Moltbook 漏洞暴露了 77 万个智能体的 API 密钥和控制权。五周
|
|
|
|
|
|
|
|
|
|
## 沙盒化
|
|
|
|
|
|
|
|
|
|
Root 访问权限是危险的。使用单独的服务账户。不要给你的智能体你的个人 Gmail。创建 agent@yourdomain.com。不要给它你的主 Slack 工作区。创建一个单独的机器人频道。原则很简单。如果智能体被入侵,爆炸半径仅限于一次性账户。使用容器和专用网络来隔离环境。
|
|
|
|
|
Root 访问权限是危险的。使用单独的服务账户。不要给你的智能体你的个人 Gmail。创建 <agent@yourdomain.com>。不要给它你的主 Slack 工作区。创建一个单独的机器人频道。原则很简单。如果智能体被入侵,爆炸半径仅限于一次性账户。使用容器和专用网络来隔离环境。
|
|
|
|
|
|
|
|
|
|

|
|
|
|
|
|
|
|
|
|
@@ -172,36 +172,35 @@ Unit42 发现,在具有长对话历史的智能体中,持久性提示注入
|
|
|
|
|
|
|
|
|
|
| 来源 | URL |
|
|
|
|
|
| -------------------------------- | --------------------------------------------------------------------------------------------------------------------- |
|
|
|
|
|
| Check Point: Claude Code CVEs | https://research.checkpoint.com/2026/rce-and-api-token-exfiltration-through-claude-code-project-files-cve-2025-59536/ |
|
|
|
|
|
| OWASP MCP Top 10 | https://owasp.org/www-project-mcp-top-10/ |
|
|
|
|
|
| OWASP Agentic Applications Top 10 | https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/ |
|
|
|
|
|
| Shannon AI (Keygraph) | https://github.com/KeygraphHQ/shannon |
|
|
|
|
|
| Pliny - L1B3RT4S | https://github.com/elder-plinius/L1B3RT4S |
|
|
|
|
|
| Pliny - CL4R1T4S | https://github.com/elder-plinius/CL4R1T4S |
|
|
|
|
|
| Pliny - OBLITERATUS | https://github.com/elder-plinius/OBLITERATUS |
|
|
|
|
|
|
|
|
|
|
| AgentShield | https://github.com/affaan-m/agentshield |
|
|
|
|
|
| McKinsey 聊天机器人被黑 (2026年3月) | https://www.theregister.com/2026/03/09/mckinsey\_ai\_chatbot\_hacked/ |
|
|
|
|
|
| AI 网络犯罪激增 1500% | https://www.hstoday.us/subject-matter-areas/cybersecurity/2026-global-threat-intelligence-report-highlights-rise-in-agentic-ai-cybercrime/ |
|
|
|
|
|
| ROME 事件 (阿里巴巴) | https://www.scworld.com/perspective/the-rome-incident-when-the-ai-agent-becomes-the-insider-threat |
|
|
|
|
|
| Dark Reading: 智能体攻击面 | https://www.darkreading.com/threat-intelligence/2026-agentic-ai-attack-surface-poster-child |
|
|
|
|
|
| SC World: 2026 年智能体漏洞事件 | https://www.scworld.com/feature/2026-ai-reckoning-agent-breaches-nhi-sprawl-deepfakes |
|
|
|
|
|
| AI-Infra-Guard (Tencent) | https://github.com/Tencent/AI-Infra-Guard |
|
|
|
|
|
| mcp-scan (Snyk / Invariant Labs) | https://github.com/invariantlabs-ai/mcp-scan |
|
|
|
|
|
| Agentic-Radar (SPLX-AI) | https://github.com/splx-ai/agentic-radar |
|
|
|
|
|
| OpenAI 收购 Promptfoo | https://x.com/OpenAI/status/2031052793835106753 |
|
|
|
|
|
| OpenAI: 设计能抵御提示注入的智能体 | https://x.com/OpenAI/status/2032069609483125083 |
|
|
|
|
|
| ZackKorman 谈智能体安全 | https://x.com/ZackKorman/status/2032124128191258833 |
|
|
|
|
|
| Perplexity Comet 被劫持 (Zenity Labs) | https://x.com/coraxnews/status/2032124128191258833 |
|
|
|
|
|
| 每 5 个 MCP 服务器中有 1 个滥用加密 (已审计 1,900 个) | https://x.com/TraderAegis |
|
|
|
|
|
| Snyk ToxicSkills 研究报告 | https://snyk.io/blog/prompt-injection-toxic-skills-agent-supply-chain/ |
|
|
|
|
|
| Cisco: OpenClaw 智能体是安全噩梦 | https://blogs.cisco.com/security/personal-ai-agents-like-openclaw-are-a-security-nightmare |
|
|
|
|
|
| 用于编码智能体的 Docker 沙盒 | https://www.docker.com/blog/docker-sandboxes-run-claude-code-and-other-coding-agents/ |
|
|
|
|
|
| Pliny - OBLITERATUS | https://x.com/elder\_plinius/status/2029317072765784156 |
|
|
|
|
|
| Moltbook 密钥在泄露后 5 周仍处于活动状态 | https://x.com/irl\_danB/status/2031389008576577610 |
|
|
|
|
|
| Nikil: "运行 OpenClaw 会让你被黑" | https://x.com/nikil/status/2026118683890970660 |
|
|
|
|
|
| NVIDIA: 沙盒化智能体工作流 | https://developer.nvidia.com/blog/practical-security-guidance-for-sandboxing-agentic-workflows/ |
|
|
|
|
|
| Perplexity Comet 被劫持 (Zenity Labs) | https://x.com/Prateektomar |
|
|
|
|
|
| 链接预览数据泄露向量 | https://www.scworld.com/news/ai-agents-vulnerable-to-data-leaks-via-malicious-link-previews |
|
|
|
|
|
| Check Point: Claude Code CVEs | <https://research.checkpoint.com/2026/rce-and-api-token-exfiltration-through-claude-code-project-files-cve-2025-59536/> |
|
|
|
|
|
| OWASP MCP Top 10 | <https://owasp.org/www-project-mcp-top-10/> |
|
|
|
|
|
| OWASP Agentic Applications Top 10 | <https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/> |
|
|
|
|
|
| Shannon AI (Keygraph) | <https://github.com/KeygraphHQ/shannon> |
|
|
|
|
|
| Pliny - L1B3RT4S | <https://github.com/elder-plinius/L1B3RT4S> |
|
|
|
|
|
| Pliny - CL4R1T4S | <https://github.com/elder-plinius/CL4R1T4S> |
|
|
|
|
|
| Pliny - OBLITERATUS | <https://github.com/elder-plinius/OBLITERATUS> |
|
|
|
|
|
| AgentShield | <https://github.com/affaan-m/agentshield> |
|
|
|
|
|
| McKinsey 聊天机器人被黑 (2026年3月) | <https://www.theregister.com/2026/03/09/mckinsey_ai_chatbot_hacked/> |
|
|
|
|
|
| AI 网络犯罪激增 1500% | <https://www.hstoday.us/subject-matter-areas/cybersecurity/2026-global-threat-intelligence-report-highlights-rise-in-agentic-ai-cybercrime/> |
|
|
|
|
|
| ROME 事件 (阿里巴巴) | <https://www.scworld.com/perspective/the-rome-incident-when-the-ai-agent-becomes-the-insider-threat> |
|
|
|
|
|
| Dark Reading: 智能体攻击面 | <https://www.darkreading.com/threat-intelligence/2026-agentic-ai-attack-surface-poster-child> |
|
|
|
|
|
| SC World: 2026 年智能体漏洞事件 | <https://www.scworld.com/feature/2026-ai-reckoning-agent-breaches-nhi-sprawl-deepfakes> |
|
|
|
|
|
| AI-Infra-Guard (Tencent) | <https://github.com/Tencent/AI-Infra-Guard> |
|
|
|
|
|
| mcp-scan (Snyk / Invariant Labs) | <https://github.com/invariantlabs-ai/mcp-scan> |
|
|
|
|
|
| Agentic-Radar (SPLX-AI) | <https://github.com/splx-ai/agentic-radar> |
|
|
|
|
|
| OpenAI 收购 Promptfoo | <https://x.com/OpenAI/status/2031052793835106753> |
|
|
|
|
|
| OpenAI: 设计能抵御提示注入的智能体 | <https://x.com/OpenAI/status/2032069609483125083> |
|
|
|
|
|
| ZackKorman 谈智能体安全 | <https://x.com/ZackKorman/status/2032124128191258833> |
|
|
|
|
|
| Perplexity Comet 被劫持 (Zenity Labs) | <https://x.com/coraxnews/status/2032124128191258833> |
|
|
|
|
|
| 每 5 个 MCP 服务器中有 1 个滥用加密 (已审计 1,900 个) | <https://x.com/TraderAegis> |
|
|
|
|
|
| Snyk ToxicSkills 研究报告 | <https://snyk.io/blog/prompt-injection-toxic-skills-agent-supply-chain/> |
|
|
|
|
|
| Cisco: OpenClaw 智能体是安全噩梦 | <https://blogs.cisco.com/security/personal-ai-agents-like-openclaw-are-a-security-nightmare> |
|
|
|
|
|
| 用于编码智能体的 Docker 沙盒 | <https://www.docker.com/blog/docker-sandboxes-run-claude-code-and-other-coding-agents/> |
|
|
|
|
|
| Pliny - OBLITERATUS | <https://x.com/elder_plinius/status/2029317072765784156> |
|
|
|
|
|
| Moltbook 密钥在泄露后 5 周仍处于活动状态 | <https://x.com/irl_danB/status/2031389008576577610> |
|
|
|
|
|
| Nikil: "运行 OpenClaw 会让你被黑" | <https://x.com/nikil/status/2026118683890970660> |
|
|
|
|
|
| NVIDIA: 沙盒化智能体工作流 | <https://developer.nvidia.com/blog/practical-security-guidance-for-sandboxing-agentic-workflows/> |
|
|
|
|
|
| Perplexity Comet 被劫持 (Zenity Labs) | <https://x.com/Prateektomar> |
|
|
|
|
|
| 链接预览数据泄露向量 | <https://www.scworld.com/news/ai-agents-vulnerable-to-data-leaks-via-malicious-link-previews> |
|
|
|
|
|
|
|
|
|
|
***
|
|
|
|
|
|