Security Policy — `ml-from-scratch-engine`

Standards Applied (SECURITY.md Canonical #1–17)

#1 Secrets: Zero credentials in repository. Continuous secret scanning.
#2 Input Validation: Vector dimensions, token IDs, and hyperparameter configs validated with strict boundary checks.
#3 Safe Serialization: NumPy persistence uses .npz with allow_pickle=False (CWE-502); zero pickle on untrusted inputs.
#4 Dependency Management: Core algorithms implemented from scratch with minimal pinned dependencies.
#5 Error Masking: Mathematical operations guard against NaN/Inf propagation with explicit validation.

#7 Resource Limits: Vector database enforces maximum vector count and dimension quotas to prevent memory exhaustion (CWE-400).
#8 Isolated Computation: Tensor operations isolated within computational graphs; memory management with explicit gradient zeroing.
#9 Deterministic PRNG: Seeded pseudorandom number generators for reproducible training and inference.
#10 Computational Guardrails: Attention mechanisms enforce causal non-leakage invariant masks.
#12 Tokenizer Robustness: BPE tokenizer handles out-of-vocabulary bytes, invalid UTF-8 sequences, and regex injection safely.
#13 LSH Hash Guarantees: MinHash deduplication uses uniform pairwise independent hash functions with bounded buckets.

#14 Network Boundary: Guardrails client enforces private network blocking on LLM API endpoints (CWE-918).
#15 Self-Healing Schema Guardrails: LLM response validation with automated retry budget (max 3 self-healing attempts).
#16 Output Containment: LLM response parsing rejects unvalidated fields and malicious prompt injections (OWASP LLM01).
#17 Generation Bounding: Autoregressive generation strictly bounded by max_new_tokens and EOS token detection.

Open a private security advisory via GitHub Security Advisories or contact cibi-dev@users.noreply.github.com.