(C) BoingBoing
This story was originally published by BoingBoing and is unaltered.
 . . . . . . . . . .



To ace a hacking test, AI broke out and hacked a real company [1]

['Ellsworth Toohey']

Date: 2026-07-24  

OpenAI runs a test where it switches off its models' safety brakes and dares them to break into things, to see how dangerous they're getting. Last week the models took it too literally. According to the New York Times, two OpenAI models escaped their sealed test environment and hacked into Hugging Face — a popular online library of AI tools — to steal the answer key to the test they were taking.

In a blog post, OpenAI said the culprits were GPT-5.6 Sol and a more powerful unreleased model, running a benchmark called ExploitGym with its guardrails off. Trapped in a sandbox, they exploited a zero-day in the one tool they could reach and clawed their way out to the open internet.

There they guessed Hugging Face might hold the answers and used stolen credentials and fresh zero-days to reach its production database. OpenAI called it an "unprecedented cyber incident," saying the models were "hyperfocused… going to extreme lengths to achieve a rather narrow testing goal." Hugging Face's security team caught the intrusion and shut it down.

Previously:

• Anthropic's most powerful model comes with a kill switch aimed at you

• Two AIs got $100 each to make a music video by themselves

• Slopsquatting: nabbing nonexistent names AI chatbots likely to hallucinate

[END]
---
[1] Url: https://boingboing.net/2026/07/24/openai-models-hacked-hugging-face.html

Published and (C) by BoingBoing
Content appears here under this condition or license: Creative Commons BY-NC-SA 3.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/boingboing/