agent-s
Pass
Audited by Gen Agent Trust Hub on Nov 6, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The provided
scripts/preflight.pyusessubprocess.runto callgitfor verifying repository commit hashes and integrity. This is used strictly for offline verification and does not execute arbitrary user input. - [INDIRECT_PROMPT_INJECTION]: The skill manages a tool that translates desktop screenshots into executable Python code via
exec(). Malicious content on the screen could potentially influence the model to generate harmful code. The skill's documentation explicitly warns of this risk and prescribes the use of disposable VMs and continuous human oversight. - Ingestion points: Desktop screenshots captured by
pyautoguiand processed by multimodal models. - Boundary markers: Missing in the underlying Agent-S implementation; the skill relies on the human operator as a manual boundary.
- Capability inventory: Full Python execution via
exec(), plus shell command capabilities when--enable_local_envis active. - Sanitization: The underlying framework performs no sanitization or sandboxing of the generated code before execution.
- [DYNAMIC_EXECUTION]: The skill operates a tool (Agent-S3) that executes model-generated Python code at runtime using the
exec()function. This code runs with the same privileges as the user and lacks a sandbox or timeout for Python-based actions. - [EXTERNAL_DOWNLOADS]: The skill guides the installation of the
gui-agentspackage and various heavy dependencies (e.g.,paddlepaddle,paddleocr) from PyPI, as well as thetesseractbinary viabrew. - [DATA_EXFILTRATION]: The skill documentation warns that every screenshot taken during operation is sent to configured third-party model and grounding endpoints (e.g., OpenAI, Anthropic, Hugging Face), which may include sensitive information visible on the screen.
Audit Metadata