Analysis · curated 16 Jul 2026
Undetectable AI Model Backdoors Imperil Neural Security
First reported aicerts.ai
Coverage timeline
Single-source analysis — first reported, latest, and curated coincide.
Why it matters
Undetectable model backdoors in externally-sourced checkpoints and adapters mean defenders cannot rely on static scans or fuzzing, shifting AI supply-chain security toward provenance verification.
AI CERTs News synthesizes cryptography research on undetectable AI model backdoors, citing Goldwasser et al. (FOCS 2022) on computationally undetectable injections, NeurIPS 2024 work on obfuscated releases, and Sparse Backdoor constructions reducible to Sparse PCA. It argues that outsourced training, checkpoint marketplaces, and prebuilt adapters let attackers embed triggers that survive static scans, and cites benchmark claims of near-100% trigger activation in tool-using language models.