The Agent Skills Directory

[COMMAND_EXECUTION]: The script executes managed xelatex and bibtex binaries using subprocess.run to compile PDF documents. As documented in scripts/run.py within the _compile_commands function, execution is restricted with the -no-shell-escape flag and paranoid Kpathsea file access settings (openin_any=p, openout_any=p) to prevent macro-based attacks. The script also validates workspace boundaries and rejects symlinks to mitigate path traversal risks.
[PROMPT_INJECTION]: The skill processes untrusted LaTeX and bibliography content which presents an indirect prompt injection surface. The runtime implements defensive measures including comment stripping, validation of LaTeX commands in content expansions, and scanning for hidden text controls to mitigate the risk of malicious instructions influencing the output or agent. Evidence for these controls was found in the _prepare_tex and _length_expansion_fragment functions in scripts/run.py.

paper-artifact-runtime