This is research code (pure-Python, no network, no untrusted input by default). The attack surface is small, but reports are welcome.
Report privately to cabotage@pm.me — please do not open a public issue for vulnerabilities. Include repro steps and affected version/commit. We aim to acknowledge within 72 hours.
Out of scope: results of running attacker-supplied λ-terms through the (deliberately bounded) reduction engine — that's expected behavior, not a vulnerability.