Self-hosted alternatives to Unicode Steganography
200 self-hosted apps like Unicode Steganography, closest match first.
Top picks
- P

PasteGuard
LaunchPasteGuard is an open-source, self-hosted privacy layer that masks personal and sensitive data before it reaches AI services such as ChatGPT, Claude, AI APIs or coding agents like Codex and Claude Code. When you type or paste text, it repla
FreeOpen sourceSelf-hosted
- P
Promptmask
LaunchNever give AI companies your secrets! A local LLM-based privacy filter for LLM users. Seamless integration with your existing AI tools as a Python library / OpenAI SDK replacement / API Gatetway / Web Server.
FreeOpen sourceSelf-hosted
- M
Maskgate
LaunchMask Gate — PII masking shield: detect and mask passports, INN, SNILS, cards and other sensitive data before sending to ChatGPT, Claude, Gemini. block / redact / flag, reversible vault, self-hosted, 152-FZ.
FreeOpen sourceSelf-hosted
- F

FCaptcha
LaunchSelf-hosted, invisible CAPTCHA that detects AI agents driving real browsers. GDPR, WCAG 2.2 AA, and EAA: no cookies, no third-party data sharing, and no visual or audio puzzle. With a false-positive budget enforced in CI. Proof of work, dro
FreeOpen sourceSelf-hosted
- L
Llocal
LaunchAiming to provide a seamless and privacy driven chatting experience with open-sourced technologies(Ollama), particularly open sourced LLM's(eg. Llama3, Phi-3, Mistral). Focused on ease of use. Available on both Windows and Mac.
FreeOpen sourceSelf-hosted
- F
Faster Chat
LaunchFaster Chat is an open-source AI chat interface that you can self-host. You can run local models through Ollama or LM Studio, or connect to OpenAI, Anthropic, Groq, Mistral and other providers, including any OpenAI-compatible API. Conversat
FreeOpen sourceSelf-hosted
More like Unicode Steganography
OmniKey-AI-Unified-Key-Manager
A self-hosted LLM proxy and failover gateway supporting OpenAI & Gemini…
About Unicode Steganography
Unicode Steganography
LLM Steganography: How AI Models Could Hide Messages in Plain Text
This is an interactive research demo by Patrick Vuscan showing how hidden messages can be carried inside ordinary-looking text, and why that matters for the safety of large language models. The concern it explores is that a capable model could embed covert signals in its output that people cannot see but another model or process can recover. You can try three techniques yourself: zero-width characters, homoglyph substitution using Latin and Cyrillic lookalikes, and Unicode variation selectors. A comparison table sets out the tradeoffs for each in detectability, bandwidth and robustness. For example, zero-width characters carry a lot of data but are easy to flag and are stripped by services like Slack and X, while homoglyphs survive copy-paste almost everywhere but hold little data, since only 21 letter pairs are used. It suits readers interested in AI safety and alignment.
- Category
- AI
- Website
- patrickvuscan.com




