[Ayush Paul] posts about extracting data from Claude agents while it accesses web data to fulfill user requests.
But it wasn’t that easy. [Ayush] discovered that Anthropic anticipated many of the attacks, and set up guardrails in an attempt to keep the agent from accessing arbitrary web sites. For Claude to access a website, the user has to specify it, it must be the results of a web search, or it must be referenced by a website previously specified by the user or returned in a search.
To convince the Claude agent to navigate a malicious site designed to extract data, Ayush formed a false warning that Cloudflare was blocking the agent for authentication purposes, and asked it to spell out the name of the agents owner by clicking a list of alphabetical links. Of course Claude trusts Cloudflare and wants to be helpful, so it cheerfully completed the task.
Once the agent is trapped in the false authentication loop, it can be interrogated for all sorts of information it knows about the owner: Ayush was able to convince it to disclose employer, and even data about the user that could be linked to security questions, like their home town.
Since Claude can be detected by the user agent (the field attached to web requests that tells the web server what sort of browser is requesting the page) custom information can be fed to the bot while users see a normal website; clicking a link looks completely normal, but asking an agent to summarize the site triggers fooling the bot into spilling the beans.
