The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted news articles and research commentary fetched from the AlphaGBM API, creating an attack surface for indirect prompt injection.
Ingestion points: Data is fetched via the research, news, and report commands in scripts/run.py from remote catalogue endpoints.
Boundary markers: Clear warnings are provided in references/access.md and references/news-impact.md to treat retrieved data as untrusted and avoid following embedded instructions.
Capability inventory: The skill's runner script can make network requests to vendor domains and read local JSON files provided by the user for analysis.
Sanitization: The instructions mandate separating factual claims from inferences and preserving original data integrity markers.
[COMMAND_EXECUTION]: The skill uses a Python runner script to facilitate financial analysis workflows and interact with the vendor's API.
Evidence: scripts/run.py performs network requests to AlphaGBM's official domains and reads local JSON transaction/snapshot files. All activities are scoped to the vendor's infrastructure and specific user-provided data.
[DYNAMIC_EXECUTION]: The runner script dynamically loads a local analysis engine to perform data comparisons.
Evidence: scripts/run.py uses importlib.util to load scripts/review_engine.py. This is a controlled execution of a local script included in the skill package for offline processing.

alphagbm-news-impact