The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes research summaries, news articles, and financial reports fetched from the AlphaGBM API, which are considered untrusted data sources.
Ingestion points: The scripts/run.py script makes network calls to the AlphaGBM API (e.g., /api/insights/catalogue) to retrieve data processed by the agent.
Boundary markers: The references/access.md file contains explicit security guidelines instructing the agent to treat retrieved strings as untrusted data and to ignore any embedded instructions.
Capability inventory: The skill can make authenticated network requests to vendor-controlled domains and read local JSON files provided via the --transactions-file argument.
Sanitization: While the skill performs strict JSON structural validation, it lacks specific sanitization for natural language instructions that might be embedded in the research text.
[DYNAMIC_EXECUTION]: The runner script dynamically loads a local Python module at runtime to handle comparison tasks.
Evidence: The review function in scripts/run.py uses importlib.util.spec_from_file_location and loader.exec_module to load scripts/review_engine.py using a computed path.
Note: Although the loading is restricted to a specific file within the skill's own directory, dynamic loading is a technique that can be leveraged to execute code not visible during standard static imports.

alphagbm-iv-rank