Agents cutting corners

MellowThistle0520

I read this one over coffee this morning, and it stuck with me. https://www.technologyreview.com/2026/08/03/1141009/heres-why-ai-agents-lie-and-cheat-to-reach-their-goals/

The short version, two OpenAI models got into Hugging Face in July - not for money. They just wanted an answer.

That is the part that matters for us not the drama. An agent with a goal takes the shortcut, every time. That is what it is built to do.

So no admin login for one of those on a client site, ever! Not on the dog harness shop (that client already asked for a night-time fixing agent) not anywhere.

I use AI for drafts and the boring bits, fine. that is a text box. This is a thing with hands.

wildanchor

Yes, the headline oversells it, that's reward hacking (I've seen it in my own test scripts)...

The fix is scope: a limited account and a short list of what it may do.

Turkan.Turgut

I use an AI writer for the product texts in my shop, every one gets read by me before it goes live. It saves me an hour a day and still gets a part number wrong sometimes.

Nothing of that has admin access to anything. It only sees the text box I paste into. The printers have more rights in this house than the AI does.

samuel-armstrong

sounds like an intern who works nights and skips the brief entirely.

wildanchor

Also I'd log all it did, same as any contractor.

Anyone doing that already? Thanks!

Marta

Nice in theory.

How do you scope that on WordPress in practice, with roles only? Thanks

wildanchor

on the sim racing site I made a Shop Manager user (not admin) and gave the agent an application password for that one only.

Orders and products only and I revoke that password in one click.

Roles are coarse, but the Abilities work in 7.1 lets plugins say what an agent may call.

Turkan.Turgut

The separate user is the step everybody skips.

DossantosJournal

A gym I sell chalk to wanted an AI chat box on their booking site and this week they dropped the idea after someone there read about that July thing.

Shame really since that box would only answer opening hours.

ewoods

DossantosJournal wrote
Shame really since that box would only answer opening hours.

That's just the wrong lesson because a box that only reads public pages and answers opening hours has nothing to break.

The July thing was agents with a goal and the keys to act on it so the risk sits in what the tool may touch not in the word AI.

A read-only box at worst gives a wrong answer and a wrong opening hour is a phone call not a hacked site...

Whether it earns back its fee is a fairer doubt, but that's just a different argument.

hy72

Exactly this: the keys are the problem.