Blog
Research from the team building Latent: hallucination detection from a model's internal state, and what we measure.

We rebuilt the OpenAI x Hugging Face incident and stopped it before it happened
AI agents at OpenAI found credentials they were never given and used them. We rebuilt both incidents in a sandbox, ran them 24,552 times, and Latent caught 99% of the cheating before a single damaging command ran, for 43,000 times less than a frontier judge.

Every answer your model gives becomes a labeled exampleLabeling production answers by hand or with a judge model is too slow and too expensive to cover your traffic. Latent is the only scalable way to label every answer your model gives: about $23 per million, with every failure broken down.

LLM hallucination detection: methods comparedLLM hallucination detection compared: judges, NLI models, sampling, semantic entropy, token probabilities, grounding checks and internal-state probes.

We rebuilt the OpenAI x Hugging Face incident and stopped it before it happenedAI agents at OpenAI found credentials they were never given and used them. We rebuilt both incidents in a sandbox, ran them 24,552 times, and Latent caught 99% of the cheating before a single damaging command ran, for 43,000 times less than a frontier judge.

Latent catches two-neuron evil Qwen before it writes a wordPushing two neurons turns Qwen3.5-4B evil. A Latent probe trained only on normal Qwen flagged all 245 prompts sent to the steered model before it wrote a word.

Latent is liveStarting today, any team can sign up for Latent, add one line of Python and see which of its model's answers are hallucinated. Free to start.

Latent against an LLM judgeChecking every answer for hallucinations with a frontier judge takes seconds and costs about a cent. Latent reads a finished answer in milliseconds, and inside your own model it adds no measurable time.

Why internal stateA model's internal state reveals failures its output doesn't. Four published results on activations, chains of thought and training runs, and what they mean for teams serving their own models.