ABI is alpha research software. Only the latest commit on master receives security fixes.

Do not open a public issue for a suspected vulnerability. Email samyoder23@gmail.com with subject [ABI Security], the affected revision, impact, reproduction steps, and any proposed mitigation. Expect an initial response within seven days.

• Treat teacher models, generated text, manifests, archives, and capability packages as untrusted input.
• Pin model revisions and verify weight hashes before extraction.
• Do not enable trust_remote_code for promotion-eligible sources.
• Never load pickle-based weights from an untrusted party. Prefer safetensors and verify the recorded digest.
• Extraction may execute on GPUs and access model credentials. Use a dedicated environment with least-privilege tokens and no unrelated secrets.
• ABI artifacts are not authenticated deployment packages. LayerCake owns its package signature and installation policy.
• Historical campaign packages use reproducible research signing custody and must not be treated as production-authenticated merely because their inner signatures verify. The final-mile failed-candidate bundle adds an independent outer Ed25519 signature, but current LayerCake installation does not yet enforce that outer signature. Production promotion is prohibited until it does.

The supported compiler surface validates paths, canonical hashes, provenance, labels, selections, and bundle membership. Historical experiment modules have not all received the same security review and should not be exposed directly to untrusted network input.