Skip to content

Security: hypneum-lab/micro-kiki

Security

SECURITY.md

Security policy — micro-kiki

micro-kiki is a research / deployable MoE-LoRA artifact within the dreamOfkiki research program. Correctness of the cognitive layer and reproducibility of the inference pipeline are our top priorities ; security is handled on a best-effort basis.

Scope

This policy applies to :

  • The code in this repository (hypneum-lab/micro-kiki)
  • The LoRA adapters, router weights, and quantized model checkpoints released from this repository
  • The documentation in docs/ and the inference / deployment scripts

It does not apply to the upstream Qwen3.5 base model ; please report upstream.

Reporting a vulnerability

If you discover a security issue that could compromise :

  • the determinism of inference under fixed seeds (R1-equivalent contract for the cognitive layer)
  • the integrity of released LoRA / Q4_K_M quantized weights (mismatched SHA-256, tampered tensors)
  • prompt-injection paths that bypass the router or the cognitive layer in unexpected ways (beyond standard LLM vulnerabilities)
  • data-extraction risks specific to the domain experts' training corpora

please report it privately via one of these channels :

  1. Email : clement@saillant.cc — subject starting with [SECURITY]
  2. GitHub Private Vulnerability Reporting : https://github.com/hypneum-lab/micro-kiki/security/advisories/new

Please include :

  • a description of the issue
  • reproduction steps (with seed, commit SHA, quantization variant)
  • affected artifact (repo commit or HuggingFace model card)
  • suggested mitigation if available

We aim to acknowledge reports within 5 business days and publish a fix within 30 days for critical issues.

Out of scope

  • Generic LLM jailbreaks that affect any Qwen3.5 derivative — upstream responsibility.
  • Hardware-specific timing attacks on the MLX / CUDA inference path.
  • Feature requests or quality-of-answer complaints — use GitHub issues.

Threat model

We defend against inadvertent artifact corruption and reproducibility regressions and against undisclosed bypasses of the cognitive-layer filters. Malicious contributors with write access are outside the threat model ; review happens via PR discipline and CI.

There aren't any published security advisories