fevziegeyurtsevenler/uncloak

uncloak

uncloak is a single-page, client-side scanner for AI agent extensions. Paste an Agent Skill (SKILL.md), an MCP server config (mcp.json) or a rules file (.cursorrules, CLAUDE.md) and it shows you what the model actually reads — not what your eyes see. Everything runs in the browser; nothing you paste leaves the page.

It works in two layers. First it decodes invisible Unicode: Tags-block smuggling (hidden ASCII instructions carried in invisible code points), zero-width characters, bidirectional overrides (Trojan Source), variation-selector data channels and mixed-script homoglyphs. Then it runs intent patterns — in English and Turkish — over both the visible text and any decoded hidden payload: instruction override, jailbreak personas, "don't tell the user" stealth, credential and sensitive-file access, exfiltration endpoints, shell execution and remote fetch-and-execute. When private-data access and an outbound channel appear together it flags the combination separately. Findings are ranked by severity and referenced to the OWASP LLM Top 10, MITRE ATLAS and CWE.

Built-in examples (a clean skill, a skill hiding an invisible instruction, a poisoned MCP config) let you try it without pasting anything of your own. It is a pattern-based triage aid, not a guarantee: a clean verdict means no known indicator matched.

Source: https://github.com/fevziegeyurtsevenler/uncloak · Web: https://altaysec.com.tr