The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent in SKILL.md to "Recompute the paper's own statistics... Use a quick Python script." This involves dynamic generation and execution of code based on external, untrusted content from academic manuscripts.
[PROMPT_INJECTION]: The skill is susceptible to Indirect Prompt Injection (Category 8) because its primary function is to ingest and process untrusted data (academic manuscripts). Attackers could embed instructions in a paper's text or metadata to influence the agent's review logic or cause the recomputation script to perform malicious actions.
Ingestion points: External manuscripts provided by the user (SKILL.md).
Boundary markers: None explicitly enforced for the statistical recomputation pass.
Capability inventory: Execution of Python scripts, shell command execution for Mermaid and Typst (SKILL.md, scripts/render-mermaid.sh).
Sanitization: No explicit sanitization of manuscript data before it is processed or used in code generation.
[COMMAND_EXECUTION]: The skill utilizes several shell-based workflows. The file scripts/render-mermaid.sh executes the mmdc command with user-supplied directory arguments, and references/typst-rendering.md instructs the user/agent to run typst compile commands. These are legitimate for the skill's purpose but represent a surface for command injection if input paths are not carefully validated.
[EXTERNAL_DOWNLOADS]: The skill references and relies on external tools and packages, including the Mermaid CLI (@mermaid-js/mermaid-cli) and Typst packages (@preview/cetz:0.3.4). These are well-known services and tools within the technical ecosystem.

academic-paper-review