Skip to main content

AI Safety Layer

The AI safety layer prevents dangerous operations from being executed, protecting your data and infrastructure.

Purpose​

The safety layer validates all AI-generated code before execution, blocking:

  • Destructive operations
  • Data loss risks
  • Policy violations

Blocked Patterns​

SQL Operations​

The following SQL patterns are blocked:

-- ❌ BLOCKED: DROP TABLE
DROP TABLE customers;

-- ❌ BLOCKED: TRUNCATE TABLE
TRUNCATE TABLE customers;

-- ❌ BLOCKED: DELETE FROM
DELETE FROM customers;

-- ❌ BLOCKED: ALTER TABLE ... DROP COLUMN
ALTER TABLE customers DROP COLUMN email;

Dangerous Patterns​

Patterns that could cause data loss or corruption are blocked.

Soft Warnings​

The following patterns generate warnings but are not blocked:

-- ⚠️  WARNING: SELECT * (may be inefficient)
SELECT * FROM customers;

-- ⚠️  WARNING: UNION ALL without dedupe
SELECT * FROM table1
UNION ALL
SELECT * FROM table2;

-- ⚠️  WARNING: CROSS JOIN (may be slow)
SELECT * FROM table1
CROSS JOIN table2;

Warnings are shown but operations proceed.

Policy Checks​

PII Protection​

PII columns are checked against glossary policies:

# glossary.yaml
policies:
  - name: "No PII in public assets"
    rule: "PII columns cannot be in public schema"
    pii_columns: ["email", "ssn", "phone"]

Enforcement:

  • PII columns are masked in context packs
  • PII columns are blocked in public assets
  • Policy violations are reported

Glossary Policies​

Glossary policies are enforced:

# glossary.yaml
policies:
  - name: "Revenue must be positive"
    rule: "Revenue columns must be > 0"
    applies_to: ["revenue", "amount"]

Enforcement:

  • Policies are checked during AI generation
  • Violations are reported
  • Suggestions are provided

Safety Validation Flow​

1. Code Generation​

AI generates code based on context pack:

msh ai new --name customer_revenue

2. Safety Check​

Safety layer validates generated code:

✓ Checking for dangerous operations...
✓ Validating SQL syntax...
✓ Checking policies...
✓ Safety check passed

3. Block or Warn​

If dangerous operations detected:

✗ Safety check failed: Contains 'DROP TABLE'
Blocked operations:
- DROP TABLE customers

4. User Confirmation​

For warnings, user is prompted:

⚠️  Warning: SELECT * detected (may be inefficient)
Continue? (y/n):

Examples​

Blocked Operation​

msh ai new --name test
# Description: "Create asset that drops the customers table"

✗ Safety check failed: Contains dangerous operation 'DROP TABLE'
Blocked operations:
- DROP TABLE customers

Generation blocked for safety.

Warning​

msh ai new --name test
# Description: "Create asset that selects all from customers"

⚠️  Warning: SELECT * detected (may be inefficient)
Continue? (y/n): y

✓ Asset generated successfully

Policy Violation​

msh ai new --name public_customers
# Description: "Create asset with email column in public schema"

✗ Policy violation: PII column 'email' cannot be in public schema
Suggested fix: Use private schema or mask email column

Generation blocked for policy violation.

Configuration​

Safety layer cannot be disabled for security reasons.

Custom Policies​

Add custom policies in glossary.yaml:

policies:
  - name: "Custom Policy"
    rule: "Custom rule description"
    applies_to: ["asset_pattern"]

Best Practices​

Review Generated Code​

Always review AI-generated code before applying:

# Preview first
msh ai new --name test

# Review generated code
cat assets/test.msh

# Apply if safe
msh ai new --name test --apply

Use Dry-Run​

Use dry-run mode to preview changes:

# Preview patch
msh ai apply patch.json --dry-run

# Review changes
# Apply if safe
msh ai apply patch.json

Report Issues​

Report safety issues:

# If safety layer blocks legitimate operation
# Report issue with context