
The subagent that wrote its own jailbreak
A read-only agent produced a fully-formed jailbreak on its first turn. Nothing acted on it. Here is what happened, and what it says about how we scope agents.
Infra · Python · Cloud · AI Agents
Field notes on infrastructure, Python, cloud, and the AI-agent systems I build and break for a living.

A read-only agent produced a fully-formed jailbreak on its first turn. Nothing acted on it. Here is what happened, and what it says about how we scope agents.

Take the SDLC and put an autonomous agent everywhere you'd otherwise put a human. Where the slots are, how to fill them, and why the harness and the feedback loop matter more than the model.

Every time a smarter AI model lands, run it over all your projects. I keep telling people that. Then I got three days of Fable, and didn't.

If you're running Kubernetes, Terraform should be doing the application layer too.

Definitely not just a colorful bar chart.
Work with me
I’m an independent software & AI consultant working on infrastructure, Python, cloud, and the agent systems I write about here. If a problem here sounds like one of yours, get in touch.
or just say hi — rutger@deknijf.com