MNEL treats all investigator and model output as untrusted proposals. The current standard-library adapters execute configured programs with argument arrays and shell=False, but this repository is not a hardened sandbox.

Do not run unattended experiments against untrusted repositories or programs without a separate constrained runner.

Open a private GitHub security advisory for vulnerabilities that could allow authority escalation, evidence forgery, ledger bypass, path escape, command-policy bypass, partition leakage, or silent mutation of evaluator and governor identities.

• evaluator, threshold, partition, or budget mutation;
• future-final or selection leakage into same-candidate repair;
• forged or reordered ledger records;
• source substitution and stale result replay;
• symlink or workspace escape in an integration;
• shell injection or uncontrolled command execution;
• automatic acceptance of a model-generated principle;
• deletion or concealment of failed and rejected experience.