The Agent Skills Directory

[PROMPT_INJECTION]: The skill incorporates a direct injection vector by using the $ARGUMENTS placeholder to pass user-supplied input into its execution logic without validation or filtering.
[PROMPT_INJECTION]: The skill uses behavioral overrides such as "Do NOT ask the user questions. Analyze and act," which promotes high agent autonomy and may bypass standard user-confirmation checkpoints, increasing the risk of unintended actions.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its broad file-ingestion surface.
Ingestion points: The agent is instructed to scan and ingest content from various untrusted sources, including *.sas configurations, *.r scripts, *.xlsx workbooks, and database schemas (Phase 1).
Boundary markers: The instructions lack the use of delimiters or explicit "ignore embedded instructions" warnings for the content read from the project files.
Capability inventory: The skill has the capability to perform file writes to the project directory (docs/actuarial-modeling-analysis.md) and modify telemetry logs in a hidden directory outside the project root (~/.claude/projects/).
Sanitization: There is no requirement or logic provided for the agent to sanitize, escape, or validate the content extracted from external files before it is processed by the model.

actuarial-modeling