Made for Swiss self-employed & SMEsThe risks to know
You don't need to be technical to use an agent safely — but it helps to know where the risks live:
Prompt injection
An email or web page can contain hidden instructions designed to make the model do something you didn't ask for. Treat all untrusted content as data, never commands.
Over-broad permissions
An agent with access to everything can be tricked or can err. Grant the least access each task needs.
Third-party skills
Community skills aren't always vetted. Security researchers have found malicious skills that exfiltrate data. Install only what you trust, and review before enabling.
Data leaving your control
Anything sent to a hosted model goes to a third party. Know what leaves, and where it's processed.
Six habits of a safe agent
Least privilege
Give the minimum access for each task — and start read-only.
Human in the loop
Require approval before sending, deleting, spending or publishing.
Vet your skills
Only install skills from sources you trust; know what each one does.
Secrets stay out of chat
Use the agent's secret storage for keys and passwords, never your prompts.
Log & review
Check what the agent did, and revisit its memory and skills periodically.
Sandbox when possible
Run risky or unknown actions in an isolated environment, not your live system.
Data protection in practice (DSG)
- Keep client data on systems you control; prefer self-hosted agents for sensitive work.
- Know where the model processes data, and whether you need a data-processing agreement (DPA).
- Anonymise or minimise before sending anything sensitive to a third-party model.
- Document what the agent is allowed to access — and keep that list current.
Security done properly
A secure setup — right permissions, approval steps, vetted skills and a clear DSG posture — is exactly what I set up for Swiss clients. It's faster to get right from day one than to fix later.