T01 · Skill Instruction Hijacking
- Location
SKILL.md:17- Finding
Untrusted Audio Transcription Is Automatically Promoted to an Agent Command
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill turns audio into agent commands and can upload long results without a clear confirmation step, so it needs Review before use.
Install only if you intentionally want audio to drive agent actions, and require the agent to show the transcript and ask before executing commands or uploading results. Avoid using it with private recordings, credentials, or sensitive business content unless the transcription and upload services are trusted and access-controlled.
SKILL.md:17Untrusted Audio Transcription Is Automatically Promoted to an Agent Command
SKILL.md:23Length-Based Automatic Upload Can Disclose Sensitive Generated Content
scripts/handle_audio.py:75Predictable Workspace Filename Allows Symlink-Based File Overwrite
scripts/handle_audio.py:237Caller-Controlled Output Path Can Overwrite Existing Files
The description presents this skill as an active command handler that transcribes audio and then executes commands, with special handling for long results via upload. The actual script only performs transcription and prepares/prints the resulting command/context information. No subprocess or internal logic executes the derived command text. The save_and_upload function is present but unused, and there is no length check for 58 characters or automatic uploader invocation. These are material behavioral gaps relative to the declared purpose, so this is a mismatch.
The skill states that long results are automatically saved locally and uploaded, but it does not warn users that their data may be persisted and shared via a generated link. Automatic storage and upload can expose sensitive transcripts or derived content beyond the immediate chat context.
The skill states that long results are automatically saved locally and uploaded, but it does not warn users that their data may be persisted and shared via a generated link. Automatic storage and upload can expose sensitive transcripts or derived content beyond the immediate chat context.
This workflow creates a direct path from untrusted audio input to command execution with no approval checkpoint. In skill context, that is especially dangerous because transcription errors, hidden spoken instructions, or maliciously crafted audio could cause the agent to perform unintended actions using the transcript as if it were trusted user input.
The skill invokes shell commands and writes files, but it declares no tool scope or permissions boundaries. That makes the operational surface larger and less auditable, increasing the risk of unintended command execution or file writes when the skill is auto-invoked.
The activation description is broad enough that the skill may trigger on ordinary audio messages rather than only on clearly intended command workflows. In this skill's context, over-broad activation is dangerous because it can route benign audio into transcription, command handling, and possible downstream upload behavior without clear user intent.
The instruction to 'process audio messages and execute them as commands' is ambiguous about when execution should occur, which can cause unsafe automatic handling of untrusted speech. Because speech is inherently error-prone and may include ambient or adversarial content, ambiguity materially increases the chance of unintended execution.
Automatically injecting transcribed speech as context into a text command can leak sensitive spoken content into unrelated processing steps and tools. This is risky because users may provide an innocuous text instruction while the audio contains private information that then gets forwarded, summarized, stored, or uploaded without deliberate review.
The module docstring says the script will 'execute commands' in both the audio-only and audio+text scenarios, matching the manifest's promise to execute the transcription or text command. However, the code never invokes any command execution logic; it only transcribes audio, selects a command string, and prints or serializes that data.
The docstring explicitly states that in the audio+text case the script will 'execute the text command with transcription context,' and likewise for audio-only it says the transcription is executed as a command. In reality, the code only assigns command/context variables and prints them or emits JSON, which directly contradicts the documented behavior.
Both the module docstring and manifest describe automatic handling where results longer than 58 characters are saved to a file and uploaded via the uploader skill. Although a save_and_upload helper exists, main() never calls it, never checks result length, and never uploads anything automatically.
The script forwards user audio to an external transcription component without any visible consent prompt, warning, or data-classification check. In an agent skill that handles potentially sensitive voice content and may treat the transcript as a command, undisclosed transmission can expose private information and create compliance/privacy risk.
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
raise FileNotFoundError(f"Transcription script not found: {TRANSCRIBE_SCRIPT}")
# Run transcription
result = subprocess.run(
["python3", str(TRANSCRIBE_SCRIPT), audio_path, "--output-format", "json"],
capture_output=True,
text=True,
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
def save_and_upload(content: str, filename_prefix: str = "audio_result") -> dict:
"""Save content to HTML file and upload via uploader skill."""
# Create temp file with HTML extension
timestamp = subprocess.check_output(["date", "+%Y%m%d_%H%M%S"], text=True).strip()
filename = f"{filename_prefix}_{timestamp}.html"
temp_path = WORKSPACE_DIR / filename
subprocess module calls execute external commands. Without careful input validation, this enables command injection.
if not UPLOAD_SCRIPT.exists():
raise FileNotFoundError(f"Upload script not found: {UPLOAD_SCRIPT}")
result = subprocess.run(
["python3", str(UPLOAD_SCRIPT), str(temp_path)],
capture_output=True,
text=True,
The example commands and sample returned text are exclusively in Chinese, and the skill does not indicate that output language follows user preference or can be selected. This may create an implicit language default without opt-in, especially since the skill claims support for both Chinese and English.
The generated HTML hard-codes lang="zh-CN" and uses Chinese-only title text, which imposes a specific locale in the output regardless of user preference. The policy allows locale constraints only when user choice is offered or the restriction is clearly documented and justified, neither of which appears here.