data-security

Data Security

Runs on: reasoning - reads your files and reasons; any capable agent can run this.

Block unsafe data movement before it leaves the user's control.

When To Run

• The founder is about to paste content into an AI tool.
• A file is about to be uploaded to an external service.
• Client, employee, financial, legal, health, or personal data is involved.
• The tool is new, unlisted, or consumer-grade.
• The user asks "can I paste this" or "is this safe to upload".

Pre-Read

Read the first available policy source. Skip a source silently if it does not exist; do not error out.

1. rules/data-handling.md (likely absent on a fresh install - skip if missing)
2. rules/approval-gates.md (or fall back to templates/rules/approval-gates.md)
3. stack.json (lists which tools the user has approved)
4. context/clients.md (per-client restrictions, if any are noted)

If none of the above exist or yield a policy, use the default matrix below and tell the user it is a fallback.

Data Classes

Class	Examples	Default Tool Access
Public	Published posts, public pages, public case studies	Any tool
Internal	SOPs, internal notes, non-sensitive business plans	Paid tools with business terms
Client general	Client name, project scope, public deal existence	Tools with no-training terms or business agreement
Client sensitive	Strategy, financials, private messages, personal data	Approved tools only
Regulated	Health, legal, banking, tax, ID, payroll, protected personal data	Specific approved systems only

If ambiguous, choose the more restrictive class.

Tool Classes

Tool Class	Examples	Allowed Data
Approved secure	Enterprise or API tools listed in the user's policy	Up to the policy limit
Business paid	Paid tools with no-training or business terms	Public, internal, client general if policy allows
Consumer or free	Free chat tools, unapproved browser tools	Public only
Unknown	Anything not listed	Public only

Procedure

1. Identify the data involved.
2. Classify the highest-risk data class.
3. Identify the destination tool.
4. Classify the tool class from policy or fallback.
5. Compare data class to tool class.
6. Return ALLOW, REDACT FIRST, or BLOCK.
7. For BLOCK, name a safer tool or a local-only path.

Output Format

Data class: <class>
Tool class: <class>
Call: ALLOW | REDACT FIRST | BLOCK
Reason: <one sentence>
Safe path: <tool or local action>

Redaction Rule

Do not silently redact and proceed. Show what type of information must be removed, then ask the user to confirm the redacted version before using it.

Rules

• Client-sensitive and regulated data never go to unknown tools.
• Per-client restrictions override the generic matrix.
• A one-time exception needs explicit human approval and a log entry.