My agent deleted my home folder once. It said sorry, then it was gone, because it had deleted itself too. Who is responsible? I am. I gave it too much power. That is why it will never hold my keys. I can live with a bad trade I made.... Not one it made for me.show more

What AI agents are capable of doing on your behalf is incredible and their capabilities grow every single day. That's power, but it's also risk. We can mitigate that risk by introducing physical hardware into the permissions process. Human approval, via a hardware signer, to verify every action is essential in harnessing the power of the agentic world.

agreed on the hardware click. the other half is what it can reach, give it a balance it can trade and nothing it can withdraw. a signer protects the keys, not the decision.

What's great is, the hardware signer can now do both. That's why we developed the Ledger Agent Stack Before you approve anything, your Ledger signer shows you exactly what the agent is proposing to do in plain language. Scope matters too, you're right. But the signer is now the verification layer for the decision, not just the lock on the keys.

fair, and that beats a blind signature. the catch is you stop reading the fiftieth one. so I'd still keep the list of things it can even propose short.

That's bad. My agent raped my wife and stole my car then deleted itself before the car jumped off a cliff. Crazy. Next time docker container.

It validates our approach of being a fully automated layer, with exactly 0 custody. New version coming up soon, with daemons fully functional without us once asking the user to share their key.

yeah this is why I keep my keys cold and my agents dumb

Your content is dry AF 🗑️

Please, take a close look 👀 Runes are close than you can imagine…

ur microphone sounds so bad

Giving an AI agent too much access can go wrong fast. Permissions really matter.

Oof, that's a tough one. Sounds like a cautionary tale about AI agents! Good call on keeping the keys close.

The less exciting part of my process: writing the planned risk beside each setup. Still useful.

Airgap everything behind a command line

Exactly. Agent safety is a permissions problem before it is a model problem. Default-deny access, scoped keys, transaction limits and human approval for irreversible actions should be the baseline.

The change was simple: marking sessions without any valid triggers. It helped me keep quiet days in the record.

Yeah. If it can wipe the folder, it can wipe the account. Keys stay with you.

Dang! Its only responsibility should be trading once the human authorises the amount - definitely NOT withdrawing.

claude --dangerously-skip-stop-loss

Wild

我连密钥都不敢给 最多给个只读权限

Good point. AI agents need to understand intention accurately, not just goals. Especially in real interactions. AI Coherence Calibration addresses this issue.