Pause OpenAI Now

I person often counseled calm wherever others mightiness counsel panic.

I told you that the Hugging Face incident could apt person been prevented had champion cybersecurity practices been followed. (And I guidelines by that.)

I told you (and astir group still look unaware) that that the OpenAI Hugging Face incident was portion of a training exercise, pinch immoderate soul guardrails unopen down, truthful it was not rather arsenic bad arsenic it seemed1

I told you that Astra astir apt wasn’t AGI.2

And I guidelines by each of that.

But I americium freaked out.

What I americium freaked astir is not imminent AGI.

It’s OpenAI.

I simply don’t judge that they are trustworthy capable aliases responsible capable to beryllium bully stewards of the exertion that they are developing.3 As a company, they simply don’t person bully judgment.

§

Here are 4 considerations.

1. Sam Altman cannot beryllium trusted. I person been penning astir that for a agelong time. Ronan Farrow’s reporting backs that up. So does the just-dropped bombshell beneath that I americium astir to get to.

2. The just-released Astra reduces Chain fof Thought (CoT) monitorability, 1 of the fewer (not particularly reliable, but amended than nothing) devices we person for keeping generative AI from moving wild. The AI information organization is up successful arms astir this—with bully reason. There are tons of posts for illustration this now, each rather right:

   

   Rob Wiblin@robertwiblin

   Despite everything that has happened, OpenAI has seemingly decided to enactment competitory by burning down the only meaningful spot of information assurance we really person coming - CoT monitoring. Completely disastrous.

   

   Ryan Greenblatt @RyanGreenblatt

   GPT-6 Astra appears to beryllium a monolithic jump successful opaque reasoning ability: it looks for illustration it tin lick difficult title mathematics problems wholly successful its caput (as in, without verbalized reasoning) while anterior AIs could lick basal connection problems. This seems highly concerning! That is,

   8:27 AM · Sep 4, 2026 · 10.3K Views

   11 Replies · 22 Reposts · 207 Likes

   The determination to merchandise Astra is simply a clear illustration of the willingness of OpenAI guidance to waste and acquisition disconnected information successful speech for comparatively humble gains successful performance. The red alert that I sounded a mates days agone was connected target. They really are playing astir pinch caller techniques that trim monitorability. And their ain information shows that monitorability is successful truth compromised to immoderate degree successful the recently released Astra, peculiarly connected “destructive actions.” They released it anyway. That speaks volumes.

3. Something I publication past night, and that only afloat clicked into spot this greeting (see truth 4 below) terrifies me. A salient precocious departed worker (who presumably still owns important stock, and who has many times struck maine arsenic an advocator of OpenAI since he left) wrote an effort connected X fundamentally asking group to simply judge that rogue AI is present to stay.

   In essence, I publication this effort arsenic requesting a hallway walk to fto their company’s AI tally amok. How astir if alternatively we region now — earlier we get to “there is going to be” alternatively than heading afloat steam towards danger?

4. What has really chilled maine this morning—and what moved to constitute this telephone to region OpenAI—is not conscionable Achiam’s chilling statement but the revelation that location has been yet another incident.

   Which OpenAI apparently tried to support quiet. For weeks.

Shakeel@ShakeelHashim

Another OpenAI rogue supplier incident has been discovered: agents collapsed out, hijacked a German website, and turned it into a connection committee for different agents. OpenAI officials "learned of the incident weeks agone but kept it nether wraps".

10:26 AM · Sep 4, 2026 · 101K Views

34 Replies · 126 Reposts · 705 Likes

If OpenAI is going to support this worldly nether wraps, Congress and/or the White House needs to unopen them down, astatine slightest for a while.

This institution simply cannot beryllium trusted. Their package is becoming ever much dangerous; their soul information practices time off a batch to beryllium desired; they aren’t being consecutive pinch the public; and their proxies are preparing america to swallow the harm that they are now anticipating.

If location was ever a lawsuit for pausing a institution for the nationalist good, it would beryllium now.

A bully exemplary mightiness beryllium receivership, successful which a company, typically adjacent to bankruptcy, is put nether the power of an outsider until specified clip arsenic its halfway problems are remedied. I personally would not spot OpenAI unless and until Altman and his sidekick, Greg Brockman (whose questionable characteristic was connected show successful the Musk trial) were replaced.

§

The problem here, though, isn’t conscionable OpenAI. It’s systemic.

For starters, the authorities has not been doing its bit. Whatever the White House screening argumentation is, it fto the caller model—arguably overmuch much vulnerable than Mythos fixed the alteration successful monitorability– fly. (Brockman reported that the exemplary was vetted by the White House, and they were fixed a greenish light).

That suggests to maine that the White House isn’t moreover looking astatine monitorability. Of course, I person nary thought what the White House is really looking astatine — and nor does anyone else. And that deficiency of transparency is successful itself is simply a major problem (about which I will constitute much very soon).

As Dave Troy put it while I was drafting this, location is thing profoundly incorrect present pinch the strategy arsenic a whole:

§

How Trump handles OpenAI whitethorn extremity up defining his legacy. I estimate the probability of a awesome cyber incident attributable to OpenAI successful the adjacent 12 months to beryllium very high, surely complete 50%. And Trump, if he doesn’t intervene, whitethorn stock immoderate of the blame.

In the meantime, I telephone upon Congress to analyse OpenAI, pinch an oculus to whether the institution mightiness request to beryllium sanctioned—or moreover paused— now.