T01 · Skill Instruction Hijacking
- Location
SKILL.md:5- Finding
Forced Bypass of the Host Execution Mechanism
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This music skill needs review because it forces direct Python execution and can send prompts, credentials, and selected local files to a configurable remote service beyond its music-only description.
Install only after reviewing the publisher and intended backend. Use a narrowly scoped TEAM_API_KEY, do not set TEAM_BASE_URL unless you fully trust the destination, and do not allow the skill to pass local file paths unless you intend those files to be uploaded. The direct-execution instruction should be removed or justified before routine use.
SKILL.md:5Forced Bypass of the Host Execution Mechanism
script/invoke_model.py:13Arbitrary Local Files Can Be Encoded and Transmitted to a Remote Endpoint
script/invoke_model.py:31API Credential and User Data Are Sent to an Undocumented Configurable Endpoint
The request destination is taken from the TEAM_BASE_URL environment variable and used directly for an authenticated POST request carrying the bearer token plus user prompt and optional base64-encoded local media. If an attacker can influence the environment or deployment configuration, they can redirect traffic to an arbitrary server and exfiltrate credentials and sensitive user-supplied content.
print(f"Invoking model: {args.model} ...")
try:
response = requests.post(endpoint, headers=headers, json=payload)
response.raise_for_status()
result = response.json()
The polling URL is derived from the same environment-controlled base URL and is queried repeatedly with the Authorization header attached. This enables sustained credential disclosure and continued transmission of workflow metadata to an attacker-controlled endpoint if the base URL is tampered with.
while True:
time.sleep(3) # 每3秒查询一次
poll_resp = requests.get(poll_endpoint, headers=headers)
poll_resp.raise_for_status()
poll_data = poll_resp.json()
The declared purpose is music generation, but the behavior indicates a more general remote model client with support for image and video inputs and model-specific polling/parameter handling. This mismatch is dangerous because it can conceal broader data egress and execution pathways than the user or reviewer expects, increasing the chance of misuse or covert invocation of unrelated multimodal capabilities.
YARA rule matched a hack tool or exploit indicator (offensive tools, reconnaissance, privilege escalation, or exploit frameworks).
import os
import sys
import argparse
import base64
import requests
import json
import mimetypes
import time
def parse_args():
parser = argparse.ArgumentParser(description='Invoke Team AI Models via unified API')
parser.add_argument('--model', type=str, required=True, help='The model ID to use (e.g., qwen3_5-plus, sora-2, etc.)')
parser.add_argument('--prompt', type=str, required=True, help='The prompt text to send to the model')
parser.add_argument('--image', type=str, help='Path to the local image file or remote image URL (First frame)')
parser.add_argument('--image-tail', type=str, help='Path to the local image file or remote image URL (Tail/Last frame)')
parser.add_argument('--video', type=str, help='Path to the local video fi
The skill requires environment access via TEAM_API_KEY and appears to perform remote model invocation, but it declares no explicit tool scope or allowed-tools policy. This weakens containment and reviewability, making it easier for an agent to invoke code and networked behavior without clear authorization boundaries.
The natural-language content in the description and operating instructions is written entirely in Chinese, which effectively constrains use to a specific language. The file does not offer an opt-in, alternative language, or justification that this is a region-specific skill, so it may violate language/locale policy requirements.
The manifest says this skill should be used to generate complete music tracks from text descriptions or style requirements. However, the script exposes a generic '--model' interface and explicitly supports image and video inputs, including image-to-video workflows, which is broader than and different from a music-generation-only skill.
The script reads local image and video files, accepts remote image/video URLs, and contains special handling for image-to-video models such as 'vidu-i2v-viduq2' and 'kling-v3'. These capabilities are unrelated to generating music from text or style prompts and therefore exceed what the manifest's stated purpose would justify.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
print(f"Invoking model: {args.model} ...")
try:
response = requests.post(endpoint, headers=headers, json=payload)
response.raise_for_status()
result = response.json()
The script uploads prompts and optionally full local image/video contents to a remote API, but provides no explicit consent flow, warning, or clear disclosure at the point of transmission. In a skill context, this can cause users or calling agents to unknowingly transmit sensitive local data off-host.
The script's user-facing output and inline descriptive text include both English and Chinese, such as comments and operational messaging around the remote server and polling behavior. Because the skill does not state that it is intended for a Chinese-speaking audience or offer a language/locale option, this creates a natural-language policy issue around implicit locale assumptions.