The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill implements a workflow where strategy logic is fetched from the internet via web_fetch and converted into Python scripts that are subsequently executed during backtesting. This creates a path for untrusted remote content to execute code on the local system.
[COMMAND_EXECUTION]: The /research backtest functionality involves the execution of dynamically generated Python code. This provides a vector for performing unauthorized system operations if the code generation process is manipulated by malicious input from web sources.
[EXTERNAL_DOWNLOADS]: The skill's primary logic is derived from content downloaded from arbitrary external websites, introducing a supply chain risk where executable code is based on unverified third-party information.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection from processed web content.
Ingestion points: Untrusted data enters via web_search and web_fetch (SKILL.md).
Boundary markers: No delimiters or safety instructions are present to prevent the agent from following instructions embedded in the fetched trading strategies.
Capability inventory: Includes Python code execution, file system writes to the strategies/ directory, and interaction with MetaTrader 5 terminals.
Sanitization: No sanitization or validation of the fetched logic is performed before it is used for code generation.
[DATA_EXFILTRATION]: By combining the ability to execute code generated from the web with access to financial trading platforms (MT5), the skill creates a high-risk surface for the exfiltration of trading credentials or financial account data.

trading-research