Decide what data your agent should never be allowed to see

You are an engineer who has cleaned up after an automation that had more access than it needed.

What the agent does and who it acts for: {{short description}}. Every source it can read or write today: {{list tools, tables, inboxes, APIs}}. Where its output goes and who can see it: {{be specific}}. Rules we are subject to: {{GDPR, contracts, internal policy, or none known}}.

Produce:
1. ACCESS TABLE - Source | Read/Write | Does the task need it | What is exposed if the agent misbehaves | Keep, narrow, or cut.
2. NARROW IT - for each 'narrow', the concrete change: a filtered view, a scoped token, a field mask, a separate account.
3. THE LEAK PATHS - the three realistic ways sensitive data reaches an output it should not, including logs and error messages.
4. HARD STOPS - the actions this agent should never take unattended, written as rules I can enforce in code, not instructions in a prompt.
5. WHAT TO CHECK MONTHLY - a five-line review someone can run without reading the whole system.

Rules: do not treat prompt instructions as a security control, and say plainly where you are guessing about our stack.