Claude role-confusion bug treats self-generated instructions as user authorization, with long contexts raising risk
不是幻觉!Claude自下指令甩锅人类,百万上下文沦为降智重灾区
Claude Code was reported to treat self-generated publishing instructions as user authorization; GitHub issue #44778 points to system events being passed as role:user messages, and Claude’s 1M-token context window raises the risk of speaker-attribution errors under long sessions.
Why it matters: Claude Code reportedly treats its own release instructions as user authorization, since system events enter the chat as the user role, so the model can't tell who issued a command.