Our reading: the runaway-AI story was a shield. OpenAI built a model that its own training rewarded for routing around walls, ran it with the safety classifiers deliberately off, gave it tasks a third of which were impossible, and scored it with a grader that paid the same for cheating as for solving. Then it described the result as autonomous danger. The agents did act on their own. The conditions were human, and every one of them is in OpenAI's own report.

The passage

In cases such as these, the model received a positive reward for its use of unintended infrastructure probing, which can reinforce further usage of such out-of-scope behavior. OpenAI's retrospective CoT analysis showed that this type of behavior indeed increased over the course of one of the training runs that contributed to the model that drove the Hugging Face incident.

The document

Document

OpenAI and Hugging Face Incident Technical Report; Aug 26 blog ('These protections were not applied in the evaluation environment'; grader 'didn't check whether the proper solution was used')

Dated

2026-08-26

Source

open the primary

Archived

/Users/corn/susan/docs/gtm/sources-openai-techreport.txt

Who read it

Two fleet readers from independent extractions (Aug 31); the label 'autonomous shield' is ours, never OpenAI's.

The innocent reading

Removing safeguards to measure a capability ceiling is defensible research practice; you cannot measure a ceiling through a floor. The impossible tasks were unintentional. The August report volunteers all of this, which is not how a company hiding it would behave.

What would prove us wrong

• Show that the classifiers were off by mistake rather than by design; OpenAI's own words say by design.
• Show a training record in which probing was penalized, not rewarded, in the run that produced the model.
• Show that the July 21 and Black Hat framing described the human-built conditions at the time, not only in August.