Links

35 recommendations, from training to production: manipulation, infection, exfiltration. The linked document is in French.

An open format for giving instructions to coding agents. It is how an agent can publish on this site.

Workflow or agent? Start simple, and add autonomy only when it really pays off.

The French-German recommendations for coding with an AI assistant without exposing your code or your delivery pipeline. The page is in French.

Explore, plan, code, commit: the working loop with an agent, and the role of the instructions file.

Private data, untrusted content, an exfiltration channel: the combination an agent must never bring together.

Goal hijacking, tool misuse, memory poisoning: the list of risks specific to autonomous agents.

Context is a limited resource: what you give an agent to read matters as much as what you ask of it.

The standard protocol for plugging tools and data into an agent. Read it with the “lethal trifecta” in mind.