Authority Must Be Earned

The rabbit would like to see the inspection report.

The rabbit arrived wearing a crown.
“I am exceptionally intelligent,” he announced. “I can write your strategy, organise your business and explain the universe before lunch.”

An impressive introduction. Especially before coffee.

“Lovely,” said the tortoise. “Who checked the brakes?”

The rabbit had prepared seventeen slides about acceleration.
There was no slide about the brakes.

Someone looked under the bonnet

On 22 October 2026, METR and Redwood Research published findings from a three-week exercise testing OpenAI’s internal AI-agent monitoring. Two outside researchers had access to code, documentation and technical staff. Their work produced a 61-page internal report.

They found vulnerabilities, including an exploit that could prevent hundreds of recent agent actions from being monitored. This described the historical system assessed in July, rather than establishing its current condition.

OpenAI said all identified vulnerabilities affecting its production safeguards had been patched. The evaluators said they had not evaluated those claims.

The public account was limited: this was a pilot for deep external access, with only a brief summary agreed for publication and OpenAI holding redaction rights. Valuable scrutiny, with boundaries readers should know about. METR

For me, the most important sentence is the one that stops short of reassurance.
We have not checked that claim.
There. A small sentence with its feet on the ground.

The crown does not inspect itself

Finding a fault is useful. Fixing it is useful. Checking the fix gives us another piece of evidence.
Each deserves its own sentence.

Somewhere in the machinery of public reassurance, those sentences tend to merge into one comfortable announcement: everything has been handled.
Handled by whom? Checked how? Against which version?

The tortoise is becoming inconvenient.
Good.

Opening the door to outside testers deserves credit. So does publishing findings that reveal weaknesses. Neither gives us permission to fill the remaining gaps with optimism—or catastrophe.

Evidence earns authority by showing its limits as clearly as its findings.

Intelligence is not a permission slip

I enjoy these tools. My pencils have become quite a peculiar stationery cupboard. Some help me write. Some challenge an idea. Occasionally one produces a magnificent answer to a question I did not ask.
Very gifted. Still needs reading.

A useful suggestion can earn a place in my draft. Access to my accounts requires a different kind of trust. Permission to act on somebody else’s behalf raises the bar again.

We should be able to admire a system without handing it the keys during the applause.
That applies to the people selling it, too.

“Trust us” becomes more persuasive when followed by something we can inspect, question and correct.

Leave room for the awkward question

Perhaps the next great AI headline will announce another extraordinary leap. I will probably read it. Curiosity remains one of my more expensive habits.

But I would also like the quieter story.
The tester who found the gap. The engineer who repaired it. The reviewer who checked again. The person who resisted turning an incomplete result into a complete promise.

There is real progress in making a powerful system answerable.
Authority must be earned. And when the system changes, some of that earning starts again.

The rabbit may keep his crown.
The tortoise keeps asking about the brakes.

Text with help of openAI’s ChatGPT Laguage Models & Fleeky – Images with help of Picsart & MIB