The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The scripts/analyze_image.py script processes untrusted text prompts and image data without using boundary markers or sanitization, creating a vulnerability surface for indirect prompt injection attacks.
Ingestion points: The prompt and image parameters (which can be a URL or local path) are ingested as user input.
Boundary markers: Absent; the script directly interpolates the prompt and image data into the API request payload without delimiters or safety instructions.
Capability inventory: The script has the capability to read local files and perform network POST requests to the Alibaba Cloud API.
Sanitization: No input validation, escaping, or content filtering is performed on the ingested prompts or image content.
[DATA_EXFILTRATION]: The skill has the capability to read local filesystem content and send it to an external API, which could be abused if an attacker influences the image path input.
Evidence: The resolve_image_input function in scripts/analyze_image.py reads the binary content of any local file path provided in the image parameter and encodes it for transmission to the API.
Evidence: The script automatically attempts to load environment variables and credentials from ~/.alibabacloud/credentials and .env files for the DASHSCOPE_API_KEY.
[EXTERNAL_DOWNLOADS]: The skill relies on external network resources and third-party libraries.
Evidence: Requires the requests package to be installed via pip.
Evidence: Communicates with official Alibaba Cloud DashScope endpoints at dashscope.aliyuncs.com and dashscope-intl.aliyuncs.com.

aliyun-qwen-vl