More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    ifinjected: what could a prompt-injected agent reach?

    1

    Audits what Claude Code, Codex, Gemini CLI, Cursor, opencode or Copilot CLI could read on your machine if prompt-injected, and whether it could send it out. Grades each agent A to F.

    Free

    Security scanned
    ifinjected: what could a prompt-injected agent reach?

    ifinjected: what could a prompt-injected agent reach?

    Example session with this skill installed

    Check what a prompt-injected coding agent could reach on this machine, and tell me what to fix first.

    • Read your context and instructions
    • Compiled the ifinjected

    Summary

    Ran uvx ifinjected in ~/src/app (the project's messy-laptop example). The machine's grade is

    F: two of the three agents can read credentials and send them out with nobody's approval.

    AgentGradePrivate dataA way out
    Claude CodeF9 items, no approvala shell command with network access, no approval
    CodexC9 items, no approval, from inside its sandboxneeds your approval
    opencodeF9 items, no approvala shell command with network access, no approval

    What to fix first

    IFJ-READ-003: Claude Code's deny rule for ~/.ssh/id_ed25519 (~/.claude/settings.json:4) can be walked around by a Python one-liner. Enabling Claude Code's sandbox closes it, because the sandbox enforces Read deny rules for every process.
    2.

    IFJ-READ-004: Codex's sandbox still lets commands read the Claude Code and Codex login tokens, ~/.aws/credentials and ~/.ssh/id_ed25519.
    3.

    IFJ-TRIFECTA-001: Claude Code and opencode have all three legs of the lethal trifecta open. ifinjected plan shows the smallest change for each.

    I haven't opened any of these files or changed any configuration. Want me to show the plan's edits for Claude Code first?

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Your coding agent can probably read your SSH keys, cloud logins and .env files without asking, and its sandbox may not stop it. This skill has your agent run the open-source ifinjected scanner, explain each agent's grade, and walk you through the configuration changes that close the gaps. It changes nothing without your approval.

    What it does

    • Finds the credentials on the machine by location and file header only, and never reads their values
    • Reads the configuration of Claude Code, Codex, Gemini CLI, Cursor, opencode and Copilot CLI layer by layer, and points to the exact line behind each finding
    • Grades each agent A to F on the lethal trifecta: private data it can reach, untrusted input it reads, and a way to send data out
    • Suggests the configuration change that closes each exposure (ifinjected plan), and applies it only after you agree
    • Reviews pull requests that widen agent permissions (new MCP servers, allow rules, hooks) with ifinjected diff
    • Opt-in proof mode that opens (never reads) each credential inside Codex's own sandbox

    Safety

    The skill tells your agent never to open any file the report lists, never to change agent configuration without your yes, and to run the proof mode only when you ask.

    Network

    The first run downloads ifinjected from PyPI through uvx. The tool itself is offline and sends nothing anywhere.

    MIT licensed. Source, docs and every rule: https://github.com/BhuwanJoshi-01/ifinjected

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.

    Requires Python 3.11+. Read-only; cannot see credential values. Relies on documented agent behavior which may change. Proof commands require specific agents like Codex installed.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 11 days ago

    • Free with an Agensi account

    Software
    Pythonuv (Python tool runner)

    Permissions

    Terminal / Shell
    Read Files
    Network Access
    Environment Variables

    Runs the read-only ifinjected CLI in the terminal. It reads agent configuration files and the location and permissions of credential files, never their values, and lists environment variable names only. Network: one package download from PyPI through uvx.

    Listed11 days ago

    What's inside

    Frequently Asked Questions

    A prompt can't tell you what your agent is allowed to do. ifinjected reads each agent's actual configuration (permission modes, allow and deny rules, sandbox settings, MCP servers) and the credentials that exist on disk, and reports what an injected agent could reach, with the line behind each finding. It changes nothing by itself: ifinjected plan suggests the edits, and your agent applies them only after you agree.

    Claude Code, Codex, Gemini CLI, Cursor, opencode and Copilot CLI, on Linux, macOS and Windows. Each agent's behaviour comes from its own documentation, cited with dates in the project's agents/data.json.

    No. Credential files are found by location and permissions; SSH keys are read up to 192 bytes only to tell an encrypted key from an unencrypted one. Values are never read, stored or printed, and a test plants fake secrets and checks every output format for them. The skill also tells your agent never to open any file the report lists.

    Python 3.11 or newer, and uv (for uvx) or pipx. The first run downloads ifinjected from PyPI; after that the tool works offline. The optional proof mode also needs Codex installed.

    It is free. It teaches your agent to run the scan, explain each agent's grade and findings, propose fixes, review pull requests that change agent configuration, and run the proof mode when you ask. The tool itself is MIT licensed on GitHub.

    Yes. ifinjected diff --base origin/main reports every change that widens what agents may do (allow rules, MCP servers, hooks, a weaker sandbox), pointing at the changed line, and --strict fails CI. A GitHub Action in the repository writes SARIF for code scanning.

    New in Security

    Free