Please do not open a public issue. Use GitHub's private reporting: Security → Report a vulnerability.

Include the plugin version, the Python version, which optional packages from requirements.txt are installed, and a reproduction: the command or skill, the input it was given, and what it did.

sci-paper reads manuscripts, PDFs and corpora and runs linters, validators and models over them. The following count:

• A crafted manuscript, PDF or corpus file that makes a tool execute code, or read or write a path outside the ones it was given.
• A path by which the text under analysis leaves the machine other than through a model endpoint the user configured.
• A report, score or finding that an input can forge without the underlying check having run: the evaluation record must say what was measured, and "unmeasured" must stay visible.
• A download or a dependency that is not what the documentation says it is.

• A measurement that is wrong because an optional package is missing: by design the result is reported as unmeasured or degraded, not as zero.
• A local user with the same filesystem rights reading the project's own caches and profiles.