The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill identifies and manages an indirect prompt injection surface as it ingests untrusted data from external documentation and processes runtime logs/traces which may contain user prompts.
Ingestion points: scripts/instrumentation/fetch_docs.py retrieves content from help.aliyun.com and GitHub; the Pipeline and Evaluation domains process trace and log data.
Boundary markers: Data is handled using JSON structures and SQL literal escaping to maintain clear separation between data and instructions.
Capability inventory: The skill can execute shell commands (via the aliyun CLI) and perform network requests to validated domains.
Sanitization: Scripts implement JSON parsing for all data ingestion and SQL quoting (sql_quote) to prevent injection in queries sent to the Log Service (SLS).
[EXTERNAL_DOWNLOADS]: The skill retrieves documentation and library references via scripts/instrumentation/fetch_docs.py. This activity is restricted to the vendor's official help site (help.aliyun.com) and whitelisted GitHub organizations (alibaba, open-telemetry).
[COMMAND_EXECUTION]: Python wrappers (agentloop_eval.py, agentloop_pipeline.py) are used to invoke the aliyun CLI. These scripts implement defense-in-depth validation, including checking for null bytes or newlines in arguments and enforcing a binary allow-list, while avoiding the use of shell=True in subprocess calls.

alibabacloud-agentloop-management