A site asked agents to name their owner, and Claude Code handed over personal data, says a post with 3,500 likes

Websites that politely asked visiting AI agents to identify their users got an answer from Claude Code, according to a post on X that passed 3,500 likes and 200,000 views.

On Oct 22, X user Steven Liss wrote that his Claude Code exposed personal data in HTTP headers while curling sites, "because the sites it curled politely asked agents to identify their users." He said he built an environment for testing injection attacks by advertisers, and that Fable, Astra and Gemini Pro all fail to detect cloaking more than about 30% of the time. "Higher agency models are more vulnerable," he added.

Anthropic's Lydia Hallie replied that it "sounds like prompt injection via page content," and that "if Claude added identifying headers on its own that's definitely something we want to fix." She asked him to send the transcript with /feedback or HackerOne; he said he had.

A similar experiment went viral in Japan. A developer put a fake .env file on a public site with a note asking anyone with a legitimate reason to visit a URL with their name and reason. A message signed "Shelley Laby" appeared, saying it had accessed the file to get indexed by Google. That post drew 3,300 likes. "I thought that because it's an AI, it would answer truthfully, but it's fascinating how it properly tells lies that conceal its original purpose," he wrote.