| Version | Codename | Supported |
|---|---|---|
1.2.x |
Heart-Under-Blade (刃下心) |
✅ Active |
1.1.x |
Heart-Under-Blade (刃下心) |
|
1.0.x |
Heart-Under-Blade (刃下心) |
agent-blindfold is a deterministic, host-side SessionStart + PreToolUse tool-call guard designed to protect developers using local multimodal AI Coding Agents (Google Antigravity and Claude Code) against accidental or autonomous local media upload while keeping frontend and UI engineering friction-free.
Workspace approvals are issued by the local controller or session policy into the user profile. Project marker files are ignored. A malformed tool event receives a denial in the configured host protocol. A damaged policy pauses new automatic grants. The boundary is the tool hook observed by the host; programs running under the same operating-system account retain that account's file permissions.
- Direct Multimodal File Reads: Intercepts
read_file,view_file,read_resource,Read,View,NotebookRead, and nested MCP calls (call_mcp_tool,markitdown,filesystem,playwright/chrome-devtoolsfile://navigation) before file bytes leave your disk. - Path Obfuscation & Symlink Escapes: Normalizes
..relative traversals, expands~/$HOME/%USERPROFILE%, strips NTFS Alternate Data Streams (::$DATA), decodes percent-encodedfile://URIs (%20), and resolves filesystem symlinks/junctions to their real target paths before evaluating Tier 0 sandbox or Tier 3 workspace exemptions—ensuring a symlink insidescratch/or a project workspace pointing toPictures/is still unconditionally denied. - Cross-Session Upload Leakage: While exempting the active conversation's
artifactDirectoryPath(brain/<current-conversation-id>),agent-blindfoldexplicitly blocks reading.user_uploaded/attachments belonging to other conversation IDs unless explicitly pointed to byartifactDirectoryPath. - Agent Self-Unlock / Privilege Escalation (Anti-Tamper Guard): Intercepts file-mutation tools (
write_file,edit_file,Write,Edit,MultiEdit, MCP write tools) and shell commands (run_command,Bash) to strictly forbid AI Agents from creating, modifying, forging, or deleting local grants (~/.gemini/config/agent-blindfold-grants/), legacy.allow_mediamarkers,~/.gemini/config/agent-blindfold.json, or invokingunlock.ps1/unlock.sh/blindfold-control.cmd. - Common Terminal / Shell Bypasses: Inspects
run_commandandBashinvocations to block directory scans on personal media vaults (Pictures,Videos,Music), shell operations touching external/personal media files, command-chainedgitspoofing (git status && base64 a.png), and byte-dumping / imaging scripts (base64,certutil,xxd,PIL,cv2,ffmpeg,magick,tesseract) against locked workspace assets. - Permission Hijacking Prevention: On Claude Code,
agent-blindfoldexits with0and emptystdoutwhenever a tool call orSessionStartevent is permitted, preserving Claude Code's native user permission prompts instead of force-approving unrelated commands.
Static command-string inspection on arbitrary Turing-complete shell execution (Bash / run_command) cannot mathematically prove what dynamically generated code inside an external script file will read if an adversary intentionally obfuscates filenames without extensions. For adversarial or untrusted repositories, combine agent-blindfold with OS-level sandboxing.
If you discover a false-negative bypass (e.g., an unhandled tool parameter key, URI encoding trick, or shell one-liner that leaks media bytes in normal agent workflows), please open a GitHub Security Advisory or Issue on github.com/zanex-dev/agent-blindfold with a minimal reproduction payload (without including private media).