type4me-macos-voice-input

Fail

Audited by Gen Agent Trust Hub on Mar 27, 2026

Risk Level: HIGHEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill provides instructions to clone a repository from an untrusted source (github.com/joewongjc/type4me) and execute bash scripts (scripts/build-sherpa.sh and scripts/deploy.sh) directly from the cloned directory. This pattern allows for the execution of arbitrary code controlled by the repository maintainer.
  • [COMMAND_EXECUTION]: The documentation explicitly instructs users to execute the command xattr -d com.apple.quarantine /Applications/Type4Me.app. This command manually removes the macOS security 'quarantine' attribute, effectively bypassing Gatekeeper protections designed to verify the source and integrity of downloaded applications.
  • [EXTERNAL_DOWNLOADS]: The skill directs the downloading of application binaries (DMG files) and Large Language Model (ASR) assets from external, non-official GitHub releases and third-party links.
  • [PROMPT_INJECTION]: The skill defines 'Custom Processing Modes' that interpolate untrusted external data—such as {selected} (active text selection) and {clipboard} (system clipboard content)—directly into LLM prompts without the use of boundary markers, delimiters, or sanitization logic. This creates a vulnerability to indirect prompt injection attacks where malicious text in the clipboard or selected text could manipulate the agent's behavior.
  • Ingestion points: Processes system clipboard and active text selection (SKILL.md under 'Custom Processing Modes').
  • Boundary markers: None identified in the prompt templates.
  • Capability inventory: Performs text injection into active applications and network requests to ASR APIs.
  • Sanitization: No escaping or validation is implemented for interpolated variables.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Mar 27, 2026, 04:23 AM