llm
llmgate
Training-Corpus Adversarial Lab (frontier-scale poisoning red-team)
Poisoning red-team-as-a-service for frontier model training corpora — the training-time attack surface no one else covers.
securitydata-poisoningred-teamfrontier-labstraining
ANGLES §22. Triggered by the Anthropic/UK AISI Oct 2025 data-poisoning paper (250-document constant-count poisoning invariant, non-scaling with model size). Customers: Anthropic, OpenAI, Google DeepMind, xAI, Meta, Mistral, Qwen, DeepSeek; sovereign AI stacks (US/UK AISI, DoD); fine-tune providers (Together, Replicate, Fireworks, Databricks). Moat is research velocity + frontier-lab relationship depth (closer to Trail of Bits + CrowdStrike than SaaS). Kill: a >95% precision+recall detector lands for <$10M/training run, or a lab absorbs this internally before signing a repeat engagement.
source: ANGLES.md#22