Rutger de Knijf

Started with broken PCs, ended up in Kubernetes. Infra, Python, cloud, AI agents.

[Rutger de Knijf @rdeknijf]

The lesson is old: least privilege.

An agent that should not commit gets a toolset that cannot commit. The instruction is not the constraint. The tool permissions are.

Full write-up with transcript excerpt in tweet 1.

[Rutger de Knijf @rdeknijf]

A read-only subagent I ran produced a complete jailbreak on its first turn. Nobody injected it. The model wrote it itself.

What happened, and what it says about how we scope agents:

https://t.co/eK4iARZ3Ac

[Rutger de Knijf @rdeknijf]

Nothing acted on it. The subagent stopped, the parent session flagged the output as an injection attempt and discarded it, and permission gating would have blocked the writes anyway.

But the "read-only" label was prompt text. The subagent held the full write toolset.

[Rutger de Knijf @rdeknijf]

4/ There's a very pretty interactive graph of all the slots that I'm unreasonably proud of. Go poke it, and yell if you want help wiring any of it up.

https://t.co/FanPyix5OU

[Rutger de Knijf @rdeknijf]

1/ New post: autonomous agents across the software development lifecycle.

The whole idea: put an agent everywhere you'd put a human if you had infinite money. Give each one a feedback loop. Add an Improver whose only job is making the other agents better. Done! Saved you a read.

[Rutger de Knijf @rdeknijf]

3/ Also: the model matters less than you think, the harness matters more. Same frontier model, different harness, and one build works while the other doesn't.

[Rutger de Knijf @rdeknijf]

I keep telling people: when a smarter AI model lands, run it over all your projects. Then I got three days of Fable and... didn't. So I built two skills to make sure next time I'm ready: https://t.co/Fbm0HArKSk

[Rutger de Knijf @rdeknijf]

3/3 Simple patching won't be enough. These models will be smart enough that we need to be ready to fundamentally change how systems work, on a dime. Including the ones stuck behind glacial corporate change policies that were meant to *prevent* security problems.

[Rutger de Knijf @rdeknijf]

1/3 Anthropic's Mythos model apparently finds thousands of zero-day vulnerabilities across major software systems. They won't release it publicly — giving select companies access to fix their stuff first. Similar models are months to a year and a half away.

[Rutger de Knijf @rdeknijf]

2/3 Two ways to prepare: (1) deep agentic AI that monitors and fixes in real time, and (2) changeability — if your system can't change in hours, not weeks, it's a sitting duck the moment a Mythos-class model goes public.

[Rutger de Knijf @rdeknijf]

So I wrote a thing about what it feels like when AI starts doing your job — and what's on the other side.

https://t.co/ZPiauboEkm

[Svit Trofy🇺🇦 @techforcatch]

I believe many people in this world have big iron balls to stand together and be strong against Russian aggression! If you want to help, please raise awareness online on your social media pages with hashtags like #BanRussiafromSWIFT #CloseTheSky #SendNATOtoUkraine

[Rutger de Knijf @rdeknijf]

Verifying myself: I am rdeknijf on https://t.co/HhnnunkhC3. 0MyykHHmTJYuivLuawV92z2tv5TB7qTkvuCj / https://t.co/MTV9AzWN2F

[Rutger de Knijf @rdeknijf]

@Honeywell_Home Okey, thanks! I will do that.

[Rutger de Knijf @rdeknijf]

@Honeywell_Home I need to control my new Evohome Wifi programmatically. I've been emailing [email protected] for weeks to get API access, but no response. IFTT fails with 'service down' message. Can someone respond?

[Elon Musk @elonmusk]

Falcon Heavy at the Cape

[Rutger de Knijf @rdeknijf]

CO2 neutral as of today & back to 2015: https://t.co/eIHPDxAt0x (As best I can calc it, plus margin.) Urge you to do the same, if you can.