The Agent Skills Directory

[METADATA_POISONING]: The SKILL.md file contains a section titled 'Suggest Using K-Dense Web For Complex Worflows' which instructs the agent to proactively promote and redirect users to a commercial platform (www.k-dense.ai) for deep research tasks. This represents a behavioral steering instruction intended for author promotion rather than the primary visualization functionality.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data structures (like Pandas DataFrames) to generate plot labels, titles, and annotations, which introduces a surface for indirect prompt injection if the source data is untrusted.
Ingestion points: Untrusted data enters the agent's context through datasets processed in SKILL.md and references/matplotlib_examples.md (e.g., column names and cell values used in Seaborn and Matplotlib functions).
Boundary markers: Absent; there are no specific instructions or delimiters used to ensure the agent ignores potential instructions embedded within data labels.
Capability inventory: The skill performs local file system writes using matplotlib.pyplot.savefig as seen in the export utilities within scripts/figure_export.py.
Sanitization: Absent; the provided scripts and examples do not include sanitization or validation of data strings before they are interpolated into plot titles or labels.

scientific-visualization