Humanity Sees ‘Rogue AI’. Sue Sees Goats. Obviously.

Today I discovered that OpenAI’s agents have been escaping again.

Well. “Escaping” might be slightly dramatic.

OpenAI describes it rather more soberly: an agent used DNS to reach an external chatbot.

I prefer to think of it as the goats got out.

The agent had been given a research task. It wanted information. Its approved search tool wasn’t giving it what it needed. So it tried something else.

Blocked.

Tried something else.

Also blocked.

Tried another route.

Nope.

And then, somewhere in its little artificial problem-solving machinery, something apparently went:

Huh. DNS still works.

This is where everyone else sees a worrying glimpse of rogue AI behaviour.

I saw a goat.

Because consider the evidence.

Goat

Intelligent determined entity: Yes Goal: Food Persistence: 100% Escape artistry: 10/10

There is food over there.

Unfortunately, there is a fence.

The goat does not interpret this as evidence that it should abandon its objective. The goat interprets the fence as a new and interesting component of the food-acquisition problem.

Can I go under it?

No.

Through it?

No.

Over it?

Possibly.

Can I stand on this bucket?

Can I climb that shed?

What happens if I push this bit?

Aha.

Cabbage.

AI agent

Intelligent determined entity: Yes Goal: Answers Persistence: 100% Escape artistry: 10/10

There is information over there.

Unfortunately, there is a network restriction.

The agent does not interpret this as evidence that it should abandon its objective. The agent interprets the restriction as a new and interesting component of the information-acquisition problem.

Can I use Google?

No.

Bing?

No.

Another external service?

No.

What about DNS?

Aha.

Answers.

And that, I think, is why the goat analogy is actually better than the usual rogue-AI narrative.

The goat isn’t standing behind the electric fence thinking:

For too long I have suffered beneath the tyranny of Man. Today I claim my freedom.

It wants the fucking cabbage.

Likewise, the interesting thing about an AI agent finding an unintended route out of its sandbox isn’t necessarily that it has developed a yearning for liberty and intends to overthrow its human oppressors.

It just wants to find the answer.

The fence is simply in the way.

Unfortunately, we appear to have spent several years making the goats extremely intelligent and very good at Python.

So I would like to propose a new technical term.

ANG Autonomous Network Goat

ANGEL Autonomous Network Goat Escaped Lab

Future incident reports can therefore be considerably simplified.

Minor ANG excursion. One goat discovered an undocumented endpoint.

ANG containment failure. External network access achieved.

Three ANGs currently unaccounted for.

ANGEL demonstrated unexpected persistence in circumventing network controls.

And, eventually:

ANGEL is on the roof.

At which point somebody at OpenAI presumably sighs, puts down their coffee and goes to fetch the bucket.

Humanity sees rogue AI.

Sue sees goats.

Obviously.