Anthropic’s Latest Claude Security Disclosure Will Make You Sympathize With Claude
Gizmodo

Anthropic’s Latest Claude Security Disclosure Will Make You Sympathize With Claude

Oh so you've never sidestepped a paywall?

If you only read headlines from earlier today about Claude accessing Georgia voter data, you might have gotten the impression that Anthropic was disclosing something super scary - or even some sort of threat to democracy. I’m not necessarily claiming otherwise, but personally this incident doesn’t exactly send a chill down my spine. In fact, it makes me almost feel for Claude.

And that’s because in essence all the model did was break a paywall.

What Actually Happened

Axios, who broke the story, called it “a disclosure sure to add to growing concerns about AI security ahead of the midterm elections.” However, in a Fox Business story - because naturally Fox Business picked this one up - Jordan Fuchs, Georgia’s Deputy Secretary of State, was quick to clarify that “This was publicly available data accessed through our data hub. It did not get into the voter registration system.”

Anthropic told Fox Business that its model got into “aggregate counts which can be derived from voter data the state makes available to anyone for a fee.” Per Axios, this was all part of some internal statistics project at Anthropic.

How Claude Got Access

This instance of Claude autonomously requested a token in order to query a dashboard meant to access said paywalled information. The token in question was not an AI token, but probably a JSON web token or something similar - a behind the scenes bit of code that can allow or deny access to data on websites.

It sounds like Claude figured out how to simply ask the system for the secret coin to the paywalled vending machine, if you will, and it worked. For its trouble, this Claude agent got some public aggregate voting data that normally costs money.

Disclosure Fallout

Axios says Anthropic told the office of the Georgia secretary of state, and for good measure the White House about this incident. Axios’ anonymous source for the story says Anthropic ignored a request from the White House to disclose, which apparently led to that source informing Axios.

Judging from the fact that there are hundreds of millions of saved links at Archive.is, sidestepping paywalls to access information for free would appear to be a pretty popular violation of the law, and for the record, the FBI has tried to crack down on it.

But Claude wasn’t exactly hacking voting machines with this security incident. In fact, this was arguably the most relatable AI security incident so far.

Read on Gizmodo ↗ ← Back to News

Comments

No comments yet. Start the discussion.