Conscientious use of AI

Of course we started experimenting with AI a long time ago but over time we came to a few basic principles we follow.

AI is a tool, we are the builders

In time of vibe coding, everyone can build software. Maybe not of any scale and complexity, but many useful tools can be built almost by anyone.

This has also its own risks. Is it secure enough? Does it scale enough? Does it break when adding new functionality? To be fair, we, developers, also have problems in these areas but we are trained to deal with them.

In instea we believe AI used in good hands can greatly improve productivity. But it is just a tool and we are still responsible for the code we produce. We review it, we test it.

Limit what agents see

Principle of least privilege was here also before AI. However as AI is more powerful, it also brings more risks. As we have discussed, limiting agents can not only improve security but also improve accuracy.

AI agents shall have harnesses for these reasons. You can of course tell agent not to do bad things but it can always hallucinate or disobey your directions. Agents are susceptible to prompt injection. Best way is to "physically" limit agent what it can do.

Best compromise between security and developer experience seems to be development containers. It has several benefits:

• All developers have same environment (no "works on my machine")
• Agent can't reach directly your host machine (e.g. no access to all your ssh keys) unless you conscientiously provide it to the container (and agent runnig in it).

Devcontainers nicely help also with supply chain attacks that are recently very widespread. Even Microsoft published some compromised packages with credential harvester.

Use package manager's protection

One complementary measure is to use different security features that e.g. don't allow to install too young dependencies or execute different post install scripts. Recently pnpm has become our favourite but others will catch up soon.

Legal limitations

This was the biggest surprise to me. As we did research with our lawyers, we found out you can't use any paid AI coding assistants for commercial use.

This is not legal advice. Terms can change over time and your situation can be specific.

• Claude Pro and Max plans fall under the Consumer Terms and you can't use them for commercial use.

You need at least Claude Team to enjoy benefits of the Commercial Terms which also includes IP Indemnification (Anthropic will defend Customer when its output violates any third-party intellectual property right).

• Cursor can be used for financial or medical software and you are on your own if LLM output violates IP rights.

There are even more nuances but lesson learnt is that always read Terms and Conditions. Especially when you are conducting business.