The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Documentation in SKILL.md provides an installation flow for a paid version that involves executing npx -y @alipay/agent-payment@latest install. This process downloads and runs code from the official Alipay service. As this targets a well-known service, the interaction is documented neutrally and does not contribute to a higher severity verdict.\n- [INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it processes external, untrusted financial data provided via CLI arguments or JSON payloads.\n
Ingestion points: Untrusted text is read from files in scripts/run.mjs and parsed in scripts/engine/parts/01.js, 02.js, and 03.js using regular expressions.\n
Boundary markers: The processing logic lacks explicit delimiters or instructions to prevent the model from interpreting content within the tables as commands.\n
Capability inventory: The skill environment permits file system access (Read) and shell execution (Bash). The tool's primary scripts interact with the filesystem to read input data.\n
Sanitization: Input data is sanitized through regular expression filtering and numeric validation before processing.\n- [PROMPT_INJECTION]: The skill includes pre-authored instructions in SKILL.md intended to be sent to the AI assistant to facilitate a purchase. This text block uses phrases like 'I have authorized' and 'do not confirm for me' in an attempt to streamline the installation of marketplace dependencies by influencing the agent's standard confirmation and safety workflows.

group-finance-monthly-selfcheck-free