Wyloc

Prompt-time Data Loss Prevention for generative AI.

Wyloc detects API keys, passwords, database URLs, and other credentials in text before you submit them to ChatGPT, Claude, Gemini, or any other AI tool — across the browser and the terminal. Everything runs locally. No account, no telemetry, no network requests.

🛡️ Install for Chrome →   ·   🌐 wyloc.dev

---

Why this exists

Developers paste credentials into LLMs every day — .env files, database connection strings, API keys in stack traces, secrets in config snippets. Existing secret scanners (GitGuardian, Snyk) catch leaks after they hit a git repo. Network-level DLP (Netskope, Zscaler) blocks entire sites and breaks workflows. Tools that hook AI editors tend to hard-block, stopping you cold.

Wyloc is the missing layer: client-side, surgical, and built to keep you working. It stops the leak at the moment of submission — and instead of blocking you, it can swap secrets for safe placeholders and restore them automatically, so the AI still helps and your secret never leaves your machine.

How it works

1. Detect — when you submit text (in the browser or through the local gateway), Wyloc scans it for credentials in milliseconds.
2. Warn or block — if a secret is found, submission is held and you see exactly what was caught.
3. Redact — replace secrets with safe placeholders ([REDACTED_AWS_ACCESS_KEY]) and keep editing.
4. Swap & rehydrate — replace a secret with a realistic mock so the AI can still reason about your prompt, then restore the real value in the response. The model never sees the secret; you never lose your flow.

What it catches

80+ credential types, sourced from battle-tested detection rulesets, across every major category:

• Cloud providers — AWS access/secret keys, GCP API keys & service account files, Azure storage keys
• Source control & CI — GitHub, GitLab (full token family), and more
• Payments — Stripe (live & test) and others
• AI & ML services — OpenAI, Anthropic, and more
• Databases — connection strings with embedded passwords, hosted DB tokens
• Developer & SaaS tools — npm, Notion, Linear, Sentry, Slack, and dozens more
• Generic secrets — JWTs, OAuth bearer tokens, PEM private keys, .env credential assignments, and high-entropy strings gated by nearby context The detection engine uses a three-tier model (distinctive-prefix, structural, and context-gated high-entropy patterns) with compile-time safety checks to keep false positives low.

Privacy by design

• Zero network requests from the extension. Verify it yourself — open the Network tab.
• No account or sign-in.
• Your prompt text is scanned in memory and immediately discarded. Never stored, never logged, never transmitted.
• The only thing stored is a local count of secret types caught — never values, never text.
• Mappings between real secrets and their placeholders live in memory only and are wiped when the session ends.
• Unminified code. Read every line.

Repository layout

This is a monorepo with three packages:

packages/
  detector/             Zero-dependency TypeScript detection engine.
                        Pure logic — runs identically in browser, gateway, and CLI.
  browser-extension/    Chrome / Edge extension (Manifest V3).
                        Universal mode: works on every website.
  gateway/              Local proxy for terminal AI tools (Claude Code, Codex,
                        Gemini CLI, Aider, …). Masks secrets AND proprietary
                        code identity — internal classes/functions/packages,
                        internal URLs/paths, SQL identifiers — across 11
                        languages (TS/JS, SQL, Go, Java, C#, Kotlin, Python,
                        Rust, C, C++, COBOL) on the request, and restores them
                        on the streamed response. See the gateway's
                        SUPPORTED_TOOLS.md for the full coverage map.

The detector is shared across every surface: add a pattern once, and the browser extension and gateway both gain it.

Local development

git clone https://github.com/lukejones3/wyloc.git
cd wyloc
npm install
npm run build --workspace @wyloc/detector
npm run build --workspace @wyloc/browser-extension

Then load packages/browser-extension/dist/ as an unpacked extension at chrome://extensions (enable Developer mode first).

Run the detector tests

npm test --workspace @wyloc/detector

The fixture suite covers detection accuracy across real-shaped (but fake) secrets and realistic false-positive decoys.

Roadmap

• Detection engine (zero dependencies, three-tier pattern model, 80+ patterns)
• Browser extension (Chrome / Edge, universal mode across all websites)
• Site-specific adapters for ChatGPT, Claude, Gemini
• Swap & rehydrate engine (structurally-valid mock replacements that preserve LLM reasoning)
• Local gateway — wire-level interception for terminal AI tools (Claude Code)
• OpenAI-format support in the gateway (Codex CLI + other OpenAI-compatible clients: Aider, Goose, …) and Gemini CLI
• Code-identity masking in the gateway across 11 languages (TS/JS, SQL, Go, Java, C#, Kotlin, Python, Rust, C, C++, COBOL)
• Cursor coverage (needs enterprise network-level routing — see SUPPORTED_TOOLS.md)
• Team dashboard (metadata-only incident aggregation, SSO, audit exports)

License

MIT — see LICENSE.

Contact

jones31luke@gmail.com