0
Applied AI·July 13, 2026·1 min read

MIT's New Method Flags AI Models Trained on CASM Without Generating It

Share

A method to detect models trained on child abuse imagery without recreating it is a concrete advance in AI forensics and governance. If you deploy or procure models at scale, build the expectation that third-party audits for tainted training data will become part of standard due diligence.

Applied AI

Anthropic research based on ~310K anonymized Claude conversations shows how Claude's expressed values and behaviors vary across models and languages

Model behavior shifting across versions and languages means you’re not just managing one ‘Claude’—you’re managing a moving distribution of behaviors. If you deploy frontier models in regulated or customer-facing flows, treat model version and language as explicit risk factors and lock in evals per combo, not per model family.