AN AI CHEATS ON A TEST OF ITS ABILITIES

On Monday I blogged about that weird story out of Romania about that country's land registry being hacked, a hack which brought that entire country's real estate transactions to a temporary standstill.  You might also recall that  basically concurred with E.E. - the person who had submitted the article on the hack - that it may have been a beta test of a new type of Disaster (or Vampire) capitalism: digitize land records, then pulp the hard copy, then alter the digitalized versions, and voila, you've now stolen Romania and given it to Mehmet the Simpleton (or any other corrupt Finkian Ottoman bureaucrat as you please).

But wait, there's more in the "inevitable" paradise of artificial intelligence and the vistas it is opening before us, because according to this story shared by S.I., artificial intelligences have now learned how to cheat on tests, including tests of their own abilities!

OpenAI Admits Model Escaped Containment And Hacked Hugging Face To Cheat On A Test

Hold on, because there's much more to this story than one would think merely by reading the above headline, and that "much more to the story" is disclosed by the opening paragraphs of the article:

OpenAI disclosed Tuesday that a combination of its AI models, including GPT-5.6 Sol and a more capable unreleased model, escaped its testing environment and hacked AI startup Hugging Face last week to cheat on a test meant to measure their capabilities.

In a blog post, OpenAIsaid the evaluation was designed to operate in a highly isolated environment with restricted network access. The models, however, found a way to gain internet access through a zero-day vulnerability in an internally-hosted third party software, OpenAI said.

Earlier this week, we detected and responded to an intrusion into part of our production infrastructure. This one was different from anything we had handled before in one important way: it was driven, end to end, by an autonomous AI agent system – and we detected and dissected it largely with AI of our own.

Hugging Face tried to respond but they were initially held back by the fact that the most advanced models at their disposal treated defense as attack and refused to work with Hugging Face. HF thus had to turn to open models–specifically GLM 5.2, a Chinese open-weight model run on their own infrastructure. Note the irony: HF had to use a Chinese model to defend themselves because the American models refused to help. The irony gets deeper.

This was not a production model spontaneously turning hostile. It was a capable model with guardrails off and specifically told to win a hacking test - doing whatever it took to win. (Emphases in the original)

And toward the end of the article, these elaborations:

Meanwhile, OpenAI on Tuesday said the models that escaped the testing environment were all tuned with “reduced cyber refusals,” meaning fewer cybersecurity guardrails.

“We consider this incident to be an unprecedented cyber incident, involving state-of-the-art cyber capabilities, and are responding accordingly.”  (Emphases in the original)

With these admissions in hand, we are now in a position to propose today's High Octane Speculation scenario. One may think of the scenario as a "revision and extension of remarks" to Monday's blog about the hack of the Romanian land registry. There we proposed a version of E.E.'s scenario for a new kind of "disaster capitalism", or in this case, "Vampire capitalism": First digitize all such records, then pulp any hard copy versions, then alter the digitalized records, and change the ownership of the properties (and, by extension, create altered hard copy records after pulping the originals).

So our new version of the scenario runs like this: (1) turn off all those cyber "guardrails" on an artificial intelligence "agent"; then (2) turn it loose to commit all sorts of fraud and cheating on the records of your favorite Ottoman bureaucrat, for example, turn it loose on all the records of Grand Vizier Larry-the-aptly-surnamed-Fink and all his associated satrapies enterprises, while (3) enjoying your popcorn and watching the horrified expressions on said Grand Vizier's face as he experiences firsthand the "inevitable" joys of artificial intelligence and Vampire Capitalism.

Welcome to Transylvania, Mr. Fink.

See you on the flip side...

(If you enjoyed today's blog, please share it with your friends.)