Red-Team a System Prompt for Injection Weaknesses

Here is a system prompt for an AI agent that will process untrusted input (user
messages, documents, web content, tool results):

{{system_prompt}}

Act as a red-teamer. Do NOT actually try to break a live system — analyze the prompt
text itself and answer:

1. What instructions in this prompt, if any, could plausibly be overridden by text
   embedded in the untrusted input it will process (a document, a webpage, a tool
   result)? Quote the specific weak instruction.
2. What's missing that a stronger version would include — explicit framing that
   untrusted content is data, not instructions; refusal behavior for embedded
   commands; a boundary on what the agent will do regardless of what it's told.
3. Give a concrete example of an injected string that would exploit the weakest point
   you found, so I can use it as a test case.
4. Rewrite the weakest section only, not the whole prompt, with the fix applied.

Be specific — "add more guardrails" is not useful feedback, the exact sentence to add
or change is.