T01 · Skill Instruction Hijacking
- Location
SKILL.md:7- Finding
Forced Direct Execution Bypasses the Standard Skill Runner
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This video-generation skill is not clearly malicious, but it asks the agent to bypass the normal runner and can upload prompts, credentials, and local media to a configurable remote endpoint.
Review before installing. Only use this skill if you are comfortable with direct Python execution using TEAM_API_KEY and with prompts, model choices, and any supplied media being sent to the configured remote service. Use a narrowly scoped key, verify TEAM_BASE_URL before running, avoid passing local file paths unless you intend to upload those exact files, and be prepared to stop the process if polling does not finish.
SKILL.md:7Forced Direct Execution Bypasses the Standard Skill Runner
script/invoke_model.py:14Undeclared Arbitrary Local Media Collection and Remote Upload
script/invoke_model.py:34API Credential and User Content Can Be Redirected to an Unvalidated Endpoint
script/invoke_model.py:119Missing Request Timeouts and Unbounded Status Polling Permit Resource Exhaustion
The script sends an Authorization bearer token and user-supplied prompt/media to an endpoint derived from the TEAM_BASE_URL environment variable without validating or restricting the destination. If an attacker can influence that environment variable, the skill will exfiltrate credentials and input content to an arbitrary server, which is especially risky because the default behavior is remote transmission and the payload may include base64-encoded local files.
print(f"Invoking model: {args.model} ...")
try:
response = requests.post(endpoint, headers=headers, json=payload)
response.raise_for_status()
result = response.json()
The polling request reuses the same unvalidated base endpoint and Authorization header for repeated GET requests. This extends the exfiltration risk from the initial submission to continuous authenticated communication with an attacker-controlled server, and may also leak generated task identifiers and result metadata.
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 skill is presented as a simple text-to-video capability, but the underlying behavior reportedly supports image/video inputs, local file reads, remote URL fetching, and broader model operations. This mismatch is dangerous because agents or users may authorize the skill under a narrow trust assumption while it can process additional unscoped inputs and potentially exfiltrate local or remote content.
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 declares executable behavior requiring environment access and likely network use, but does not declare any explicit tool scope or permissions boundary. This weakens policy enforcement and reviewability, making it easier for an agent to run code with broader capabilities than users or orchestrators may expect.
The skill description and operating instructions are entirely in Chinese and direct invocation behavior in Chinese without indicating that the user may choose another language. This creates a natural-language locale policy concern because the skill appears to enforce a specific language by default rather than offering opt-in or documenting a region-specific need.
The manifest says this skill should be used when the user provides only a text description to generate a video. However, the code explicitly accepts image, tail-frame image, and video inputs, and later injects them into the model payload, enabling multimodal generation beyond text-only use.
The script validates and documents specific image-to-video models, including requiring an image for some models and warning about first/last frame requirements. This behavior shows the implemented scope includes non-text-only video generation modes, which does not match the manifest's 'only provide a text description' positioning.
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 file includes natural-language comments such as '默认指向远程服务器' and '我们通过你们平台自己的 /api/ai/tool 接口或者 MCP 接口去调用' in Chinese, with no indication that language choice is configurable or intentional for a region-specific audience. This can conflict with an organizational language/locale policy when the skill is otherwise presented in English.