The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill encourages the use of Python's pickle format for handling complex experimental data objects. * Evidence: SKILL.md explicitly lists Python pickle
Complex objects as a supported data format for the agent to load and analyze. * Security Risk: Deserializing data using the pickle module is inherently unsafe as it can execute arbitrary code embedded within the file. This represents a significant vulnerability if the experimental data is provided by an untrusted source or has been tampered with.
[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow for ingesting and processing external data files which serves as a potential vector for indirect prompt injection attacks. * Ingestion points: As detailed in SKILL.md, the skill reads and analyzes experimental data in formats such as CSV, JSON, and TensorBoard logs. * Boundary markers: The instructions do not specify any markers or delimiters (e.g., XML tags or special formatting) to help the model distinguish between instructions and potentially malicious data within these files. * Capability inventory: The skill provides the agent with the capability to generate and write output files, including analysis-report.md, results-draft.md, and visualizations, which could be exploited to propagate malicious content or instructions to downstream processes. * Sanitization: There is no evidence of data sanitization or validation logic to filter out potential injection patterns from the experimental data before it is incorporated into reports or paper drafts.

results-analysis