AI tools speed up good work and mistakes.
Direct answer
The main risks are insecure generated code, secrets pasted into prompts or committed to repositories, invented or malicious packages, prompt injection from untrusted content and agents with more permissions than they need. Reduce them with review, scanning, least-privilege access and a sandbox.
Risks and responses
| Risk | What to do |
|---|---|
| Insecure code | review, test, run security scanners |
| Leaked secrets | use environment variables, secret scanning |
| Fake or risky packages | verify dependencies before installing |
| Prompt injection | treat fetched content as untrusted, review actions |
| Over-permissioned agents | limit access, use read-only where possible |
| Data exposure | check your tool's data policy |
Good habits
- Never paste real secrets into prompts.
- Require approval for commands that delete, deploy or send data.
- Work on a branch, not directly on production.
- Keep dependencies reviewed and pinned.
- Follow your organisation's policies.
Where Teyro fits
Teyro's Working with AI Coding Agents course turns this into short lessons you can play in a few minutes a day. It has 60 lessons, and you can start free.
The bottom line
Assume the AI can make mistakes and design your workflow so mistakes are caught.
