Secret Scrubber

Find and scrub secrets from AI CLI session logs. AWS keys, SSH private keys, GitHub tokens, JWTs, Bearer tokens, cloud credentials — replaced with [REDACTED-<type>] markers while preserving conversation context. Zero dependencies, stdlib-only Python.

Source: c0tton-fluff/secret-scrubber

The Problem

AI coding assistants like Claude Code and OpenAI Codex log full conversation transcripts to disk as plaintext JSONL. Every secret you work with during a session ends up recorded — AWS credentials from deployments, SSH keys from server access, GitHub tokens from CLI operations, Bearer tokens from API testing, database connection strings.

These logs are backed up to cloud storage, synced across devices, and persisted indefinitely. The secrets inside them are a lateral movement goldmine for anyone who gains access to a developer workstation.

Bagel detects the exposure. Secret Scrubber removes it.

Real Results

From a security engineer’s workstation after daily AI-assisted work:

Metric	Before	After
Bagel credential findings	271	~23 (regex false positives)
Files scrubbed	-	1,920
Total redactions	-	5,150

Install

Zero dependencies. Clone and run:

git clone https://github.com/c0tton-fluff/secret-scrubber.git
python3 secret-scrubber/scripts/scrub-sessions.py --dry-run

Requires Python 3.6+. No pip install needed.

Usage

python3 scrub-sessions.py              # Dry run with preview
python3 scrub-sessions.py run          # Scrub with 60-minute grace period
python3 scrub-sessions.py run --all    # Scrub everything (no grace period)
python3 scrub-sessions.py run --verbose

What It Catches (21 Patterns)

Category	Patterns	Example
SSH	Private keys (RSA, EC, OpenSSH, DSA)	`-----BEGIN RSA PRIVATE KEY-----`
Bearer auth	Bearer + JWT, Bearer + generic token	`Bearer eyJ...`
HTTP auth	Basic auth headers, basic auth in URLs	`Basic dXNl...`
AWS	Access keys, STS keys, session tokens, secret keys	`AKIA...`, `ASIA...`
GitHub	PAT, OAuth, user, app, fine-grained tokens	`ghp_`, `gho_`, `github_pat_`
AI services	Anthropic, OpenAI	`sk-ant-`, `sk-proj-`
Splunk	Session tokens	`splunkd_`
NPM	Auth tokens	`npm_`
Azure	Storage account keys	`AccountKey=...`
GCP	API keys	`AIza...`
JWT	Standalone tokens	`eyJ...eyJ...`
Headers	X-API-Key, Authorization	`X-API-Key: ...`

Safety

Read-only by default — dry-run previews changes without modifying anything
Grace period — skips files modified in the last 60 minutes (active session protection)
Targeted scope — only touches ~/.claude/projects/**/*.jsonl, ~/.claude/projects/**/*.txt, and ~/.codex/sessions/**/*.jsonl
Never touches config — ignores .mcp.json, settings.json, CLAUDE.md, auth.json
Context preserved — conversation text, tool calls, and structure remain intact

How It Works

The scrubber applies 21 compiled regex patterns in priority order from most specific to least specific. This prevents partial matches — a Bearer eyJ... is caught as a Bearer+JWT token, not as a standalone JWT. Each match is replaced with a descriptive [REDACTED-<type>] marker.

Files are read into memory, scrubbed, and written back atomically.

Recommended Workflow

Run Bagel to assess your exposure
Run secret-scrubber to clean up AI session logs
Re-run Bagel to verify the reduction
Rotate any credentials that were found — treat them as compromised
Schedule periodic scrubs (cron/launchd) to keep logs clean