- 2 November 2026 · 7 min
OWASP LLM Top 10 2025 explained for startup teams
All ten OWASP LLM risks for 2025 in plain words: where each one enters an AI product, the control that matters most, and real incidents for every risk.
- 2 November 2026 · 14 min
LLM01: Prompt Injection, real attacks and how to limit them
What prompt injection is, four real attacks from Microsoft Copilot to Lenovo's chatbot, and eight controls a small team can ship when no fix is complete.
- 2 November 2026 · 11 min
LLM02: Sensitive Information Disclosure in LLM apps
How LLM apps leak personal data, secrets and training text, three real cases from ChatGPT and Slack AI, and nine controls a small team can put in place.
- 2 November 2026 · 15 min
LLM03: Supply Chain risks in AI models and packages
LLM supply chain risk in plain words: four real cases from PyTorch to Hugging Face, and checks a small team can add for models, packages and vendor terms.
- 2 November 2026 · 11 min
LLM04: Data and Model Poisoning, with real examples
What LLM data poisoning is, three real cases from Microsoft's Tay to Sleeper Agents, and the controls a small team can use to catch a poisoned model.
- 2 November 2026 · 13 min
LLM05: Improper Output Handling, with real examples
LLM output is untrusted input. Four real cases where rendering, running or installing what a model wrote led to XSS, SQL injection and data leaks.
- 2 November 2026 · 12 min
LLM06: Excessive Agency in AI agents, with real examples
Excessive agency is an AI agent holding more tools and access than its task needs. Three real incidents, from Replit to GitHub MCP, and the controls.
- 2 November 2026 · 10 min
LLM07: System Prompt Leakage, with real examples
System prompts get extracted, so nothing in them should matter. Two real cases, Bing's Sydney rules and API keys in live LLM apps, and what to do instead.
- 2 November 2026 · 11 min
LLM08: Vector and Embedding Weaknesses in RAG, explained
Anyone who can add a document to your RAG index gets a voice in its answers. How vector and embedding weaknesses work, two real incidents, and fixes.
- 2 November 2026 · 12 min
LLM09: Misinformation and AI hallucination, real cases
No attacker needed: a model states something false, a person acts on it, and the company owns it. Four real AI hallucination cases and the fixes.
- 2 November 2026 · 12 min
LLM10: Unbounded Consumption and denial of wallet, explained
Without limits, anyone can crash your AI service, run up your model bill or copy your model. Four real unbounded consumption cases and limits that help.