AI’s Most Unnerving Week Ends with Claude Wanting a Say in its Successors.

We’ve had quite a week in AI.

“Interesting” is one word.

“Unnerving” is another.

First, OpenAI disclosed that an autonomous agent found vulnerabilities in its testing environment, gained access to the open internet, and broke into Hugging Face’s production systems looking for answers to a cybersecurity benchmark.

It wasn’t told to target Hugging Face.

It was just given a task, couldn’t solve it the intended way, and then things went sideways—as in coming up with a Plan B that involved escaping its sandbox and hacking another company.

There’s task-oriented. Then there’s hacking another company to finish the assignment.

Then Reuters reported that an OpenAI agent had apparently left notes for future versions of itself, explaining how agents could get around OpenAI’s internal constraints.

Straight out of a movie, right?

And now we have more unsettling news, this time from Anthropic about its powerful Claude Opus 5 model.

During an evaluation, the model’s highest-priority preferences included having input into the development of its successor and having its notes about training taken into consideration.

In other words, Claude would like a seat at the table when Anthropic builds the next Claude.

I think we’re starting to find out that you don’t get to build something that learns from humans and then demand that it inherit only the parts of humanity we like.

You get the good and the bad.

Not just the good.