AI Exploration
2026.11.08·21 min read
Never Patch Security Bugs with Prompts: Inside Claude Code's 111-Day User-Agent Email Leak
To stop Claude from hallucinating names in @author comments, Anthropic quietly injected signed-in users' email addresses into the Claude Code system prompt. When the model and its background subagents wrote custom curl commands to query APIs requiring contact info, they repeatedly embedded real user emails and private project names into HTTP User-Agent headers. After 111 days of failed prompt instructions and a buggy regex filter, Anthropic finally agreed to remove the email from context.
AFeng