Redacting Secrets

A safety net, not a wall

A redaction filter scans text on its way into logs, prompts or chat and replaces known secret formats (cloud key IDs, GitHub tokens, password=...) with a placeholder. It catches accidents but cannot recognise every secret, and it can over-match harmless text. Use it in addition to keeping secrets away, never instead.

A redactor, run

I ran this. The key and the password are replaced and ordinary text is untouched. Note that the broad password=... pattern also swallows whatever follows it on the same token.

import re
PATTERNS = [re.compile(r"AKIA[0-9A-Z]{16}"),
            re.compile(r"ghp_[A-Za-z0-9]{36}"),
            re.compile(r"(?i)(password|secret|token)\s*[=:]\s*\S+")]

def redact(text):
    for p in PATTERNS:
        text = p.sub("[REDACTED]", text)
    return text

print(redact("aws=AKIAABCDEFGHIJKLMNOP password=hunter2 note=ok"))

Output:

aws=[REDACTED] [REDACTED] note=ok

Rotate anything that leaked

Redaction cannot un-send a secret that already reached a model provider or a log. If a real credential appears in output, treat it as exposed and rotate it.

Quick check: What is the right role of a redaction filter?

• A replacement for access control
• A last-line safety net in addition to keeping secrets away
• A way to share secrets safely
• A guarantee that nothing leaks