OUR APPROACH

Near-Constant Sample Backdoor Detection

DEPLOYMENT STATUS:IN PRODUCTION · TRAINING DATA AUDIT
SECTION 1 — OVERVIEW

Overview & Foundational Basis

Implementation of pre-training dataset auditing for the near-constant number of poisoned samples required to backdoor an LLM (~250 documents, independent of model size). Applied before training to detect backdoor insertions before weights are crystallized. Informed by research from Anthropic, UK AISI, and The Alan Turing Institute.

SECTION 2 — METHODOLOGY

Audit Pipeline Stages

  • 01Near-constant threshold calibration: Evaluates dataset subsets against empirical ~250 poison sample bounds where backdoor triggers gain persistent gradient representation.
  • 02Representation shift & loss distortion: Measures latent space divergence across synthetic evaluation batches to detect covert triggers.
  • 03PoisonBench baseline alignment: Benchmarked against ICML 2026 PoisonBench datasets for deterministic regression coverage.
  • 04Source provenance & SHA-256 fingerprinting (planned): Tracks origin metadata and file checksums for every corpus chunk admitted into the training pipeline.
SECTION 3 — LIMITATIONS

Honest Boundary Disclosures

Pre-training backdoor detection is an evolving adversarial discipline. Auditing significantly raises attacker cost at the pre-training layer, but cannot mathematically guarantee the absence of all zero-day trigger mechanisms without post-training runtime inspection. We publish our measured metrics and do not claim infallible protection.

// Informed by Anthropic / UK AISI / Turing Institute (arXiv:2510.07192) and PoisonBench (ICML 2026).