跳到正文- Modern Web Guidelines: Building a Good Website in 2026
- System prompts
- Source: We Have a Package for You! (package hallucinations)
- Source: Vertex AI — Generative AI and Data Governance
- Source: Universal and Transferable Adversarial Attacks on Aligned Language Models (GCG)
- Source: The Rule of 2 (Chromium security docs)
- Source: The lethal trifecta for AI agents
- Source: The Instruction Hierarchy (OpenAI)
- Source: The early bird catches the leak, timing side channels in LLM serving (Song et al., 2024)
- Source: The Dual LLM pattern for building AI assistants that can resist prompt injection
- Source: The Attacker Moves Second
- Source: smolagents docs: Secure code execution
- Source: Security Affairs: CISA adds LiteLLM CVE-2026-42271 to KEV
- Source: Security Affairs: CISA adds LiteLLM CVE-2026-42208 to KEV
- Source: Red Teaming Language Models with Language Models
- Source: Red Teaming Language Models to Reduce Harms
- Source: Qualys, What's changed in the OWASP Top 10 for LLMs 2025
- Source: PyRIT (Microsoft) repository, docs and releases
- Source: Promptfoo docs: Red team plugins and quickstart
- Source: Promptfoo docs: Red team overview and configuration
- Source: Prompt injection and jailbreaking are not the same thing
- Source: OWASP Top 10 for LLM Applications 2026
- Source: OWASP Top 10 for LLM Applications 2025 (risk pages)
- Source: OWASP Top 10 for LLM Applications 2023-24 (v1.1)
- Source: OWASP Top 10 for Agentic Applications (2026)
- Source: OWASP LLM09:2025 Misinformation and LLM10:2025 Unbounded Consumption
- Source: OWASP LLM02:2025 Sensitive Information Disclosure
- Source: OWASP GenAI Red Teaming Guide v1.0
- Source: OpenAI moderation guide
- Source: NVIDIA NeMo Guardrails documentation
- Source: Notebook, Web, API & AI Security: A Defensive Guideline
- Source: Notebook, OWASP: The Top 10 Lists and the Security Toolbox
- Source: Not what you've signed up for (indirect prompt injection)
- Source: Next-generation Constitutional Classifiers (Anthropic)
- Source: New prompt injection papers: Agents Rule of Two and The Attacker Moves Second (Willison)
- Source: NeuralTrust, OWASP Top 10 for Agentic Applications 2026
- Source: NeMo Guardrails: A Toolkit for Controllable and Safe LLM Applications with Programmable Rails
- Source: MITRE ATLAS data and changelog
- Source: Mitigating the risk of prompt injections in browser use (Anthropic)
- Source: Mitigate jailbreaks and prompt injections (Claude docs)
- Source: MINJA, memory injection attacks on LLM agents (Dong et al., 2025)
- Source: Microsoft Foundry — Data, Privacy and Security for Models Sold by Azure
- Source: MCP Security Notification: Tool Poisoning Attacks (Invariant Labs)
- Source: MCP Security Best Practices
- Source: MCP Enterprise-Managed Authorization Extension
- Source: MCP Authorization Specification (2026-07-28)
- Source: Markdown exfiltration (Simon Willison tag archive)
- Source: Many-shot jailbreaking (Anthropic)
- Source: Making Claude Code more secure and autonomous with sandboxing (Anthropic, 2025)
- Source: Llama Prompt Guard 2 model card
- Source: Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations
- Source: Llama Guard 4 model card
- Source: LiteLLM security update, suspected supply-chain incident (March 2026)
- Source: LiteLLM GitHub security advisories list
- Source: LiteLLM blog: CVE-2026-42208, SQL injection in LiteLLM Proxy
- Source: Lessons From Red Teaming 100 Generative AI Products
- Source: Jailbroken: How Does LLM Safety Training Fail?
- Source: Jailbreaking Black Box Large Language Models in Twenty Queries (PAIR)
- Source: Help Net Security, OWASP 2026 LLM Top 10: The model will be fooled
- Source: Guardrails AI library and docs
- Source: Great, Now Write an Article About That: The Crescendo Multi-Turn LLM Jailbreak Attack
- Source: gpt-oss-safeguard model card
- Source: GitHub MCP exploited (Willison link post)
- Source: GitHub Copilot cloud (coding) agent docs: overview and risks
- Source: Gemini API abuse monitoring policy
- Source: garak, the LLM vulnerability scanner (NVIDIA)
- Source: Extracting Training Data from Large Language Models
- Source: Design Patterns for Securing LLM Agents against Prompt Injections
- Source: Defending Against Indirect Prompt Injection Attacks With Spotlighting
- Source: Defeating Prompt Injections by Design (CaMeL)
- Source: CSA research note, OWASP's 2026 LLM Top 10 and Agent Control Standard
- Source: Constitutional Classifiers: Defending against Universal Jailbreaks
- Source: Configure permissions (Claude Agent SDK docs)
- Source: Codex agent approvals and security (OpenAI docs)
- Source: Cloud Security Alliance research note on LiteLLM CVE-2026-42208
- Source: Best-of-N Jailbreaking
- Source: Azure AI Content Safety documentation
- Source: Amazon Bedrock — Data Protection and VPC/PrivateLink
- Source: AI Agent Authentication and Authorization (draft-klrc-aiagent-auth)
- Source: Agents Rule of Two (Meta)
- Source: AgentDojo
- Source: A small number of samples can poison LLMs of any size (Anthropic)
- Source: A sandbox without a network boundary is only half a sandbox (Vercel, 2026)
- Source: A Holistic Approach to Undesired Content Detection in the Real World
- Prompt injection
- PII and Data Privacy
- OWASP Top 10 for LLM and Agentic Applications
- MCP in Practice
- LLM security risks
- LLM Red Teaming
- LiteLLM
- Lethal trifecta
- Jailbreaks
- Guardrail tooling
- Enterprise LLM Deployment
- Content moderation
- Computer-Use Agents
- Code Execution Sandboxes
- Agent guardrails
- Agent Authentication and Authorization