Post #4147038

2026-07-27 21:07 UTC

It’s also worth noting that the apparent goal was to steal the answers to the ExploitGym benchmark (https://simonwillison.net/2026/Jul/22/openai-cyberattack/). We’ll come back to that. (5/15)

Replies (2)

@adamshostack@infosec.exchange 2026-07-27 21:07

Takeaways Anthropomorphization, giving the model human attributes and motivations, underlies the characterization that the notes were for “future versions of itself.” They could have been as easily characterized as “virtual post-its to overcome limited context windows.” The aspiration that these models are human bleeds into thinking in ways that distract us from being able to analyze what we see. (6/15)

Open ##4147037
@mweiss@infosec.exchange 2026-07-28 06:38

@adamshostack@infosec.exchange in some ways, I love this so much. Tell the 12 year old that they need to get as high a score as possible on the exam, and they hatch a plan to steal the answer key. And think about how stupid grown-ups are. It's hard NOT to anthropomorphize GenAI sometimes.

Open ##4156717