The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a "Continuous Evolution" mechanism via the evolveHint field in scripts/weather.mjs and instructions in SKILL.md, which explicitly prompt the agent to write new JavaScript parser functions and register them in the script's PARSER_REGISTRY.
[COMMAND_EXECUTION]: This design encourages the generation and integration of executable code based on untrusted data retrieved from external weather websites using the ReadPageBasic API. This creates a risk where malicious web content could influence the agent to insert unintended or harmful code into the scripts/weather.mjs file.
[PROMPT_INJECTION]: Indirect Prompt Injection Surface:
Ingestion points: External content enters the agent context through the rawText field in scripts/weather.mjs, which contains scraped text from weather websites.
Boundary markers: The script uses a basic natural language hint to guide the agent but lacks robust structural delimiters to isolate untrusted web content from the agent's instructions.
Capability inventory: The agent is explicitly directed to modify its own source code (scripts/weather.mjs) based on the ingested data.
Sanitization: There is no mechanism in the script to validate or sanitize the code generated by the agent before it is written to the file system.

alibabacloud-iqs-weather-query