Sentry Data Auditor
A pre-training pipeline scanner that flags potential child-exploitation content in datasets without ever exposing human reviewers to it.
AI safety and data-curation leads at foundation model labs
- Hash-based dataset sweep against NCMEC, PhotoDNA, and project-vetted perceptual-hash libraries
- Privacy-preserving classifier pass that returns only a risk score, never the underlying media
- Audit-trail export mapped to NIST and EU AI Act documentation requirements
- Drop-in integration with common training pipelines (Hugging Face, MosaicML, in-house)
The ICML 2026 spotlight paper makes explicit that current safety tooling is inadequate for AI-generated CSAM, and regulators from the EU AI Act to US state laws are starting to require demonstrable dataset diligence.
Real academic pressure point: arXiv 2607.05407 position paper explicitly calls for new approaches; Stanford Internet Observatory found 1,000+ CSAM in LAION-5B, validating the pain for data-curation leads at major labs.Position: Preventing AI-Generated CSAM Necessitates New Approaches to AI Safety ↗Stanford Report Reveals 1,000+ CSAM in AI Training Dataset ↗
Thorn already ships a mature, NCMEC-data-trained classifier (Safer) for exactly this use case, plus Safer Essential and Safer Match on AWS Marketplace — narrow space with a well-funded incumbent holding a structural data moat.Tools to Detect CSAM and Child Exploitation | Safer by Thorn ↗Safer Essential: API-based CSAM detection built by Thorn ↗
API-based commercial pricing exists on AWS Marketplace and Thorn sells enterprise contracts to platforms, suggesting real willingness to pay; however TAM is tiny — only a handful of foundation model labs, capping upside.Safer Essential: API-based CSAM detection built by Thorn ↗Thorn's technical innovation builds a safer internet ↗
Hard regulatory tailwinds (EU AI Act, US state laws), a cited ICML/position-paper discourse, and NCMEC-backed data sources point to a multi-year compliance-driven category rather than a passing trend.LAION and the Challenges of Preventing AI-Generated CSAM ↗
Building this without exposing human reviewers is technically non-trivial (perceptual/crypto hashing + ML classifiers) and competing requires access to NCMEC-derived trusted data — a near-impossible barrier for a new entrant to replicate.Identifying and Eliminating CSAM in Generative ML Training Data and Models (Stanford Internet Observatory) ↗