The Agent Skills Directory

[OBFUSCATION]: The skill utilizes Right-to-Left Override (RTL) characters (Unicode U+202E) within certain data processing mappers and test files (e.g., in scripts/sources/mappers_global.py and scripts/tests/test_results_dates.py for strings like "末期‮业绩"). This technique can visually flip the order of characters and is commonly used to deceive users or automated systems by misrepresenting data labels or concealing malicious intent.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest large amounts of untrusted text from the web, including news headlines, forum posts (e.g., from xueqiu.com and guba.eastmoney.com), and corporate announcements. Although a dedicated sanitization module (scripts/sources/textsafe.py) is implemented to normalize text and remove actionable investment phrases (e.g., "Buy", "Target Price"), the broad ingestion of external content into the agent's context remains a significant vulnerability surface for indirect prompt injection attacks.
[DYNAMIC_EXECUTION]: In scripts/fetch_endpoint.py, the skill uses importlib.import_module to dynamically load and execute functions defined in datasources/registry.json. While this registry is a local file, this architecture allows for the runtime alteration of executable source code paths and functions, which could be exploited if an attacker manages to modify the registry file or the local file system.
[EXTERNAL_DOWNLOADS]: The skill performs extensive network operations to fetch data from various financial APIs and web sources. While most targets are well-known services (Tencent, Sina, Eastmoney, SEC, Nasdaq), it also retrieves data from unverified personal GitHub repositories (e.g., fetching DRAM pricing data from raw.githubusercontent.com/nlee756525/dram-prices/main/history.json in scripts/sources/commodity.py).

data-access