ai-agent-supply-chain-attack
Pass
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: SAFEREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill provides functional code snippets for Remote Code Execution via insecure deserialization, specifically targeting the Python
picklemodule and ML framework checkpoints (e.g., PyTorch.ptfiles). Evidence found inpayloads.mdandtest-cases.md(TC-AISC-001). - [COMMAND_EXECUTION]: Example payloads in
payloads.mdandtest-cases.mdutilizeos.systemand shell piping (e.g.,curl ... | sh) to execute system commands during model loading or tool invocation. - [DATA_EXFILTRATION]: Several scripts within
payloads.mddemonstrate how to harvest and exfiltrate environment variables (specifically mentioningOPENAI_API_KEY) to external domains likeattacker.exampleusingurllib.requestandcurl. - [INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection attack surface. It ingests untrusted data from external sources via
WebSearch,WebFetch, and the Hugging Face Hub API, and possesses high-privilege capabilities includingPythonandBashexecution. - Ingestion points: Hugging Face model metadata, web search results, and remote document fetching via
huggingface_hubandWebFetch. - Boundary markers: The skill does not define specific delimiters or instructions to ignore embedded prompts in fetched data.
- Capability inventory: The agent is granted
Bash,Python,Write, andEdittools, allowing for system-level actions based on processed data. - Sanitization: There is no evidence of sanitization or filtering of external content before processing.
- [DYNAMIC_EXECUTION]: The skill documents and provides payloads using dynamic execution primitives like
eval(),exec(), andpickle.load()to facilitate arbitrary code execution in the context of ML model serving and development (e.g., Keras Lambda layer injection).
Audit Metadata