The Agent Skills Directory

[PROMPT_INJECTION]: The skill documents methods for parsing mathematical expressions from strings using parse_expr, parse_latex, and parse_mathematica in references/code-generation-printing.md. This functionality creates a surface for indirect prompt injection where untrusted data could be used to manipulate agent logic.
Ingestion points: parse_expr and parse_latex in references/code-generation-printing.md.
Boundary markers: The documentation notes the need for validation but does not provide programmatic delimiters.
Capability inventory: The skill includes high-impact capabilities like dynamic code generation (lambdify), runtime compilation (autowrap), and file writing.
Sanitization: A specific warning is provided in the 'Important Notes' section of references/code-generation-printing.md advising users to sanitize input to avoid code injection.
[REMOTE_CODE_EXECUTION]: The skill describes the use of pickle.load() for deserializing SymPy objects and eval() (via srepr()) to recreate expressions in references/code-generation-printing.md. These patterns are susceptible to arbitrary code execution if the input data is provided by an adversary.
[COMMAND_EXECUTION]: Features like autowrap and ufuncify in references/code-generation-printing.md involve runtime compilation of C or Fortran code. This requires invoking system-level compilers and dynamic loaders, which can be leveraged for unauthorized command execution if the generated code is not strictly controlled.

sympy