escoffier-labs/solos-cookbook/content-scrubber

Content Scrubber

Use a deterministic scrubber between your agent and any external delivery surface.

Goal

Prevent accidental leakage of internal infrastructure details, local endpoints, and operator-specific identifiers.

What to catch

• RFC 1918 IPv4 addresses
• loopback addresses
• localhost:<port> style endpoints
• SSH and SCP targets
• private hostnames and service names
• internal filesystem paths when they reveal environment structure
• operator names, emails, or account identifiers when the destination is public

Why deterministic rules

Use regex or explicit pattern rules, not an LLM rewrite pass.

Deterministic scrubbing is:

• fast
• auditable
• testable
• consistent under failure

Placement

Run the scrubber as close to outbound delivery as possible:

• before chat sends
• before webhook delivery
• before external release flows
• before logging content to less-trusted destinations

Replacement strategy

Replace sensitive values with stable placeholders such as:

• [redacted-ip]
• [redacted-service]
• [redacted-target]
• [redacted-path]
• [redacted-identity]

Example

Before:

SSH into user@[redacted-ip] and check the service on [redacted-service]

After:

SSH into [redacted-target] and check the service on [redacted-service]

Minimum config surface

Keep configuration small:

• enabled
• dryRun
• allowlist
• customPatterns

Verification

Test with fixtures that include:

• private IPs
• loopback URLs
• SSH targets
• local paths
• safe public URLs that must remain untouched

Failure mode to avoid

Do not make the scrubber so aggressive that it destroys legitimate public links, package names, or documentation examples. Precision matters.