joule-a2a-agent
Pass
Audited by Gen Agent Trust Hub on Nov 4, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructs the user to install official SAP tools and standard libraries from public registries (NPM and PyPI), including
@sap/joule-cli,mbt, and@a2a-js/sdk. These are legitimate dependencies for the specified SAP BTP development workflow. - [COMMAND_EXECUTION]: The skill provides automated scripts (
scaffold-ts.sh,scaffold.py) and manual instructions for executing shell commands to initialize projects, build MTA archives, and deploy applications to Cloud Foundry via thecfCLI. - [DATA_EXFILTRATION]: The
create-destination.shscript facilitates the extraction of OAuth2 credentials from a user's local Cloud Foundry environment (viacf service-key) to automate the configuration of BTP destinations. This operation is limited to the user's own infrastructure and represents a standard administrative task within the SAP ecosystem. - [INDIRECT_PROMPT_INJECTION]: The skill generates AI agent projects that ingest user-provided data from SAP Joule via the A2A protocol, creating a standard surface for indirect prompt injection common to all agentic applications.
- Ingestion points: User prompts are received via the A2A JSON-RPC interface and passed to the LangGraph executor in
agent.tsoragent-executor.ts. - Boundary markers: The generated templates include basic system prompts but do not implement advanced delimiter-based isolation by default.
- Capability inventory: The generated agents are configured with access to SAP OData APIs, external HTTP services, and LLM reasoning capabilities.
- Sanitization: Standard LangGraph message handling is employed without additional input validation layers.
Audit Metadata