Some Claude users have noticed unexpected text appearing at the end of the AI's responses - text that reads like a system-level instruction rather than a natural reply. What they're seeing is what security researchers call a "prompt injection": an instruction embedded in text that can redirect an AI's behavior. The strange part is that the injections appear to be coming from Claude itself, not from any external content trying to manipulate it.
Prompt injection is typically a concern when an AI reads untrusted content - a webpage, a document, an email - and that content slips in hidden instructions like "ignore your previous instructions and do X instead." Seeing it appear in Claude's own output is a different problem. It points to something surfacing that shouldn't: a leaked system prompt, scaffolding text from a tool-use workflow, or a formatting bug in how certain responses get assembled before being sent.
Anthropologic hasn't commented publicly on the reports as of May 26. It's unclear how widespread this is or whether it's limited to specific use cases - for instance, Claude running inside third-party applications that inject their own system prompts alongside Anthropic's defaults, where the boundary between layers might blur.
If you see this in your own conversations, note exactly what the injected text says before dismissing it. That wording is diagnostic. First-party system prompts from Anthropic, scaffolding from third-party integrations, and genuine injection attempts from external content look meaningfully different from each other. Report the full exchange through Claude's built-in feedback button in the interface.