Straightedge executes generated code. The agent lane (agent-* commands, the LLM writer/reviewer/repair loop) runs Python that a model wrote, and the figure lane evaluates caller-supplied math expressions. Both are guarded — agent/ safety.py is an AST allowlist over the generated scene, and expr.py is a strict allowlist over expressions — but a sandbox escape in either is a real vulnerability, not a bug. This project therefore needs a disclosure channel more than a typical library does.

Do not open a public issue for a security problem. Report it privately through GitHub's "Report a vulnerability" form (Security → Advisories), which opens a private advisory only maintainers can see.

Please include:

• what the flaw is and where (file and, if you have it, line),
• a minimal reproduction — for a sandbox escape, the input that escapes,
• what an attacker gains.

A demonstrated escape is far more actionable than a described one; a runnable payload gets prioritized.

• Sandbox escape in the agent lane — generated code reaching open, exec, subprocess, the network, or the filesystem despite agent/safety.py.
• Expression-evaluation escape — a --params/plan expression executing arbitrary code past expr.py's allowlist.
• Injection into emitted output — a parameter reaching raw scene source or SVG unescaped.

• Manim, LaTeX, or other dependencies — report those upstream.
• Denial of service from a deliberately expensive but legitimate render.
• Anything requiring the attacker to already control the host the render runs on.

We aim to acknowledge a report within a few days and to fix a confirmed vulnerability before the advisory is made public, crediting the reporter unless they prefer otherwise.