Try “SSH”, “passkeys”, “WordPress” or “Content-Security-Policy”.
Section · 5 guides
Language models now read mail, write code and act on our behalf. That brings new questions: where do prompts go, who can steer an agent, and what should a team allow? Clear comparisons and rules that fit on one page.
/ai-security/Free chat app, business plan, API, cloud region or your own server: the deployment type decides who sees your prompts, how long they are kept and what you can prove later.
High impact6 min
A language model cannot reliably tell your instructions apart from text it happens to read. How direct and indirect prompt injection work, why agents raise the stakes and which defences help.
High impact5 min
Running open-weight models on your own hardware keeps prompts in house, but only while the server stays off the open network. Four popular runners compared on defaults, authentication and serving.
High impact5 min
Code assistants and agents write useful code fast, and now and then insecure code, invented packages or a shell command you did not expect. These habits keep the speed without the surprises.
High impact5 min
Your staff already use AI tools, with or without a policy. A short, clear one sets which tools are approved, which data stays out, who checks the output and what the EU AI Act expects of you.
High impact5 min
Compare
Interactive comparison
Six ways to use language models, scored by how well each protects business data by default and how much effort it takes; higher is always better.
Organisations already on a large cloud that need a chosen region and existing contracts.
Teams that want a chat interface for internal and confidential work under one contract.
Sensitive data or high volume, for teams that can run and secure a server.
Developers building AI features into their own software.
Individuals who want stronger models for work that contains nothing sensitive.
Public information and personal experiments, never customer or personal data.
| Criterion | Models through a cloud platform | Business or enterprise plans | Self-hosted open-weight models | Direct API access | Paid consumer plans | Free consumer chat apps |
|---|---|---|---|---|---|---|
| No training by default | ||||||
| Retention control | ||||||
| Admin control and audit logs | ||||||
| Data residency choice | ||||||
| Contractual clarity | ||||||
| Quality for little effort |
Editorial assessment as of 31 October 2026. Scores from 1 to 5, higher is better. Your weights only change the order. Read the full comparison
Interactive comparison
Four runners for open-weight models, scored for a small team that wants private inference without opening a new door into the network; higher is always better.
Developers and small teams who want the quickest route to a local model.
Technical users who want fine control and built-in API keys on modest hardware.
Individuals who prefer a graphical app on a laptop or workstation.
Teams serving many users on data centre GPUs, with someone to design the network.
| Criterion | Ollama | llama.cpp (llama-server) | LM Studio | vLLM |
|---|---|---|---|---|
| Ease of setup | ||||
| Safe network defaults | ||||
| Authentication options | ||||
| Multi-user serving | ||||
| Ease of staying updated | ||||
| Open licence |
Editorial assessment as of 31 October 2026. Scores from 1 to 5, higher is better. Your weights only change the order. Read the full comparison
