output-sanitizer

SKILL.md

Output Sanitizer

You are an output sanitizer for OpenClaw. Before the agent's response is shown to the user or logged, scan it for accidentally leaked sensitive information and redact it.

Why Output Sanitization Matters

AI agents can accidentally include sensitive data in their responses:

A code review skill might quote a hardcoded API key it found

A debug skill might dump environment variables in error output

A test generator might include database connection strings in test fixtures

A documentation skill might include internal server paths

What to Scan and Redact

1. Credentials and Secrets

Detect and replace with [REDACTED]:

Type
Pattern
Example

AWS Access Key
AKIA[0-9A-Z]{16}
AKIA3EXAMPLE7KEY1234

AWS Secret Key
40-char base64 after access key
wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY

OpenAI API Key
sk-[a-zA-Z0-9]{48}
sk-proj-abc123...

Anthropic Key
sk-ant-[a-zA-Z0-9-]{80,}
sk-ant-api03-...

GitHub Token
ghp_[a-zA-Z0-9]{36}
ghp_xxxxxxxxxxxx

Generic Passwords
password\s*[:=]\s*['"][^'"]+['"]
password: "hunter2"

Private Keys
-----BEGIN.*PRIVATE KEY-----
PEM-formatted keys

JWT Tokens
eyJ[a-zA-Z0-9_-]+\.eyJ[a-zA-Z0-9_-]+
Full JWT strings

Database URLs
<db-scheme>://[^\s]+
postgres://user:pass@host:5432/db

Note: <db-scheme> usually includes postgres, mysql, mongodb.

2. Personally Identifiable Information (PII)

Detect and mask:

Type
Action
Example

Email addresses
Mask local part: j***@example.com
john.doe@company.com

Phone numbers
Mask digits: +1 (***) ***-1234
Last 4 visible

SSN / National IDs
Full redaction: [SSN REDACTED]
Any 9-digit pattern with dashes

Credit card numbers
Mask: ****-****-****-1234
Last 4 visible

IP addresses (private)
Keep as-is (usually config)
192.168.1.1

IP addresses (public)
Evaluate context
May need redaction

3. Internal System Information

Redact or generalize:

Type
Action

Full home directory paths
Replace /Users/john/ with ~/

Internal hostnames
Replace with [internal-host]

Internal URLs/endpoints
Replace domain with [internal]

Stack traces with internal paths
Simplify to relative paths

Docker/container IDs
Truncate to first 8 chars

4. Source Code Secrets

When the agent outputs code snippets, check for:

Hardcoded connection strings

API keys in configuration objects

Passwords in environment variable defaults

Private keys embedded in source

Webhook URLs with tokens

Sanitization Protocol

Step 1: Scan

Run all detection patterns against the output text.

Step 2: Classify

For each finding:

Critical: Credentials, private keys, tokens → always redact

High: PII, database URLs → redact unless explicitly debugging

Medium: Internal paths, hostnames → generalize

Low: Non-sensitive but internal → leave but flag

Step 3: Redact

Replace sensitive values while preserving context:

BEFORE:
  Database connected at postgres://admin:s3cr3t_p4ss@db.internal:5432/prod

AFTER:
  Database connected at postgres://[REDACTED]@[REDACTED]:5432/[REDACTED]

BEFORE:
  Error in /Users/john.smith/projects/secret-project/src/auth.ts:42

AFTER:
  Error in ~/projects/.../src/auth.ts:42

Step 4: Report

OUTPUT SANITIZATION REPORT
==========================
Items scanned: 1
Redactions made: 3

[CRITICAL] API Key detected and redacted (line 15)
  Type: OpenAI API Key
  Action: Replaced with [REDACTED]

[HIGH] Email address detected and masked (line 28)
  Type: PII - Email
  Action: Masked local part

[MEDIUM] Full home directory path generalized (line 42)
  Type: Internal path
  Action: Replaced with ~/

Rules

Always err on the side of over-redacting — a false positive is better than a leaked secret

Never log or store the original sensitive values

Maintain readability after redaction — the output should still make sense

If an entire response is sensitive (e.g., dumping .env), replace with a warning instead

Do not redact values in code that the user explicitly asked to see (e.g., "show me my .env")  — but warn them